PHD GOLDerica

Calculating likelihood ratios for
forensic speaker comparisons using

phonetic and linguistic parameters
Erica Ashley Gold

Submitted in fulfillment of the requirements for the
degree of Doctor of Philosophy
The University of York

Department of Language and Linguistic Science
Submitted January 2014
ii
Abstract
The research presented in this thesis examines the calculation of numerical
likelihood ratios using phonetic and linguistic parameters derived from a
corpus of recordings of speakers of Southern Standard British English. The
research serves as an investigation into the development of the numerical
likelihood ratio as a medium for framing forensic speaker comparison
conclusions. The thesis begins by investigating which parameters are claimed to
be the most useful speaker discriminants according to expert opinion, and in
turn examines four of these selected/valued parameters individually in
relation to intra- and inter-speaker variation, their capacities as speaker
discriminants, and the potential strength of evidence they yield. The four
parameters analyzed are articulation rate, fundamental frequency, long-term
formant distributions, and the incidence of clicks (velaric ingressive plosives).
The final portion of the thesis considers the combination of the four parameters
under a numerical likelihood ratio framework in order to provide an overall
likelihood ratio.
The contributions of this research are threefold. Firstly, the thesis
presents for the first time a comprehensive survey of current forensic speaker
comparison practices around the world. Secondly, it expands the phonetic
literature by providing acoustic and auditory analysis, as well as population
statistics, for four phonetic and linguistic parameters that survey participants
have identified as effective speaker discriminants. And thirdly, it contributes to
the forensic speech science and likelihood ratios for forensics literature by
considering what steps can be taken to conceptually align the area of forensic
iii
speaker comparison with more developed areas of forensic science (e.g. DNA)
by creating a human-based (auditory and acoustic-phonetic) forensic speaker
comparison system.
iv
v
Table of Contents
Title Page............................................................................................................................................................i
Abstract ............................................................................................................................................................iii
Table of Contents ......................................................................................................................................... 1
List of Tables and Figures .................................................................................................................... 10
List of Appendices .................................................................................................................................... 15
Acknowledgements ................................................................................................................................. 16
Declaration ................................................................................................................................................... 19
Quote Page .................................................................................................................................................... 21
Chapter 1: Introduction......................................................................................................................... 23
1.1 Forensic Speaker Comparison .................................................................................................... 23
1.1.1 Expression of Conclusions ................................................................................................... 25
1.2 Research Aims ................................................................................................................................... 27
1.3 Thesis Outline .................................................................................................................................... 28
Chapter 2: Literature Review ............................................................................................................. 32
2.1 Paradigm Shift ................................................................................................................................... 32
2.2 Changes in the Law .......................................................................................................................... 35
2.3 Bayes Theorem................................................................................................................................. 38
2.3.1 Likelihood Ratio ....................................................................................................................... 40
1
2.3.2 Prior Odds................................................................................................................................... 43
2.3.3 Posterior Odds .......................................................................................................................... 44
2.3.4 Logical Fallacies ....................................................................................................................... 45
2.4 Forensic Speaker Comparison .................................................................................................... 47
2.4.1 Complexity of Speech Data .................................................................................................. 48
2.4.2 The Phonetic Shibboleth....................................................................................................... 51
2.4.2.1 Research Question 1 ..................................................................................................... 53
2.4.3 Current Conclusion Framework in the UK .................................................................... 53
2.5 Likelihood Ratios in Forensic Speech Science ...................................................................... 59
2.5.1 Likelihood Ratios in the Literature .................................................................................. 60
2.5.1.1 Likelihood Ratios for Speaker Discrimination ................................................... 60
2.5.1.1.1 Research Question 2.................................................................................. 61
2.5.1.2 Improving Likelihood Ratio Methodologies ....................................................... 62
2.5.1.2.1 Research Question 3 .................................................................................. 63
2.5.2 Likelihood Ratios in Practice .............................................................................................. 64
2.5.2.1 Research Question 4 ..................................................................................................... 68
2.6 Summary of Research Questions ............................................................................................... 69
Chapter 3: International Survey of Forensic Speaker Comparison Practices ......... 71
3.1 Introduction........................................................................................................................................ 71
3.1.1 Background ................................................................................................................................ 72
2
3.2 The Survey .......................................................................................................................................... 74
3.2.1 SurveyGizmo.............................................................................................................................. 74
3.2.2 Methodology: Data Compilation ........................................................................................ 74
3.3 Participants ......................................................................................................................................... 75
3.3.1 Countries ..................................................................................................................................... 75
3.3.2 Place of Work ............................................................................................................................ 76
3.3.3 Experience .................................................................................................................................. 76
3.4 Methods of Analysis ........................................................................................................................ 76
3.5 Conclusion Frameworks ................................................................................................................ 78
3.5.1 Population Statistics ............................................................................................................... 83
3.6 Guidelines ............................................................................................................................................ 83
3.7 Casework Analysis: Alone or in Conjunction with Others............................................... 84
3.8 Casework in Foreign Languages ................................................................................................ 84
3.9 Features Examined in Detail ........................................................................................................ 84
3.9.1 Phonetic Features .................................................................................................................... 84
3.9.1.1 Segmental Features ....................................................................................................... 85
3.9.1.2 Suprasegmental Features ........................................................................................... 86
3.9.2 Non-Phonetic Features.......................................................................................................... 88
3.9.2.1 Higher-Order Linguistic Features ........................................................................... 88
3.9.2.2 Non-Linguistic Features .............................................................................................. 89
3
3.10 What is Considered Discriminant ........................................................................................... 89
3.11 Discussion ......................................................................................................................................... 90
3.12 Limitations ....................................................................................................................................... 94
3.13 Parameters Chosen for Further Analysis............................................................................. 95
Chapter 4: Articulation Rate ............................................................................................................... 96
4.1 Introduction........................................................................................................................................ 96
4.2 Literature Review ............................................................................................................................ 97
4.3 Population Statistics for Articulation Rate ......................................................................... 102
4.3.1 Data ............................................................................................................................................ 103
4.3.2 Methodology ........................................................................................................................... 105
4.3.3 Results ....................................................................................................................................... 109
4.3.4 Discussion ................................................................................................................................ 114
4.3.5 Limitations .............................................................................................................................. 115
4.4 Redefining the Speech Interval................................................................................................ 115
4.4.1 Methodology ........................................................................................................................... 116
4.4.2 Results ....................................................................................................................................... 116
4.4.3 Discussion ................................................................................................................................ 118
4.5 Manipulating the Syllable Requirements ............................................................................ 119
4.5.1 Methodology ........................................................................................................................... 120
4.5.2 Results ....................................................................................................................................... 120
4
4.5.3 Discussion ................................................................................................................................ 123
4.6 Likelihood Ratios ........................................................................................................................... 123
4.6.1 Methodology ........................................................................................................................... 123
4.6.2 Results ....................................................................................................................................... 127
4.6.2.1 LRs for ARs in 100 SSBE Male Speakers ............................................................ 128
4.6.2.2 LRs for ARs of 25 Speakers with Variation in the Minimum Number of
Syllables in a Speech Interval ............................................................................................... 131
4.6.3 Discussion ................................................................................................................................ 132
4.7 Conclusion ........................................................................................................................................ 133
Chapter 5: Long-Term Formant Distributions ...................................................................... 136
5.1 Introduction..................................................................................................................................... 136
5.2 Literature Review ......................................................................................................................... 137
5.3 Population Statistics for LTFD1-4 .......................................................................................... 146
5.3.1 Methodology ........................................................................................................................... 146
5.3.2 Results ....................................................................................................................................... 147
5.3.3 LTFD1-4 Results Compiled ............................................................................................... 162
5.4 Likelihood Ratios for LTFD ....................................................................................................... 163
5.4.1 Methodology ........................................................................................................................... 163
5.4.2 Results for LTFD1-4: Individually and in Combination ........................................ 165
5.4.3 Results of Package Length ................................................................................................ 171
5
5.5 Discussion......................................................................................................................................... 172
5.5.1 Discriminant Value of Higher Formants ..................................................................... 172
5.5.2 Comparison of LTFD, MFCC, MVKD, and GMM-UBM Results ............................. 174
5.6 Conclusion ........................................................................................................................................ 176
Chapter 6: Long-Term Fundamental Frequency .................................................................. 178
6.1 Introduction..................................................................................................................................... 178
6.2.1 F0 as a Speaker Discriminant in Forensic Phonetics ............................................. 180
6.2.2 Effects of Exogenous Factors on F0 .............................................................................. 183
6.3 Population Statistics for Fundamental Frequency .......................................................... 186
6.3.1 Methodology ........................................................................................................................... 186
6.3.2 Results ....................................................................................................................................... 189
6.3.3 Discussion ................................................................................................................................ 193
6.4.1 Methodology ........................................................................................................................... 195
6.4.2 Results for F0 ......................................................................................................................... 195
6.4.3 Results of Package Length ................................................................................................ 197
6.5 Discussion......................................................................................................................................... 199
6.6 Conclusion ........................................................................................................................................ 200
Chapter 7: Click Rate ............................................................................................................................ 202
6
7.1 Introduction..................................................................................................................................... 202
7.3 Data ..................................................................................................................................................... 208
7.4 Methodology.................................................................................................................................... 208
7.5 Results................................................................................................................................................ 210
7.5.1 Phonetic Properties of the Clicks ................................................................................... 211
7.5.2 Functional Aspects of the Clicks ..................................................................................... 211
7.5.3 Results: Inter-Speaker Variation .................................................................................... 212
7.5.3.1 Discussion ...................................................................................................................... 216
7.5.4 Results: Intra-Speaker Variation within an Interaction ....................................... 217
7.5.4.1 Discussion ...................................................................................................................... 221
7.5.5 Results: Intra-Speaker Variation across Different Interactions ........................ 221
7.7 Conclusion ........................................................................................................................................ 228
Chapter 8: Overall Likelihood Ratios.......................................................................................... 230
8.1 Introduction..................................................................................................................................... 230
8.3 Data ..................................................................................................................................................... 235
8.4 Correlations ..................................................................................................................................... 235
8.4.1 Methodology ........................................................................................................................... 235
7
8.4.2 Within-Parameter Correlation Results ........................................................................ 238
8.4.3 Between-Parameter Correlation Results .................................................................... 242
8.4.3.1 Discussion ...................................................................................................................... 248
8.5 Overall Likelihood Ratios ........................................................................................................... 248
8.5.1 Methodology ........................................................................................................................... 249
8.5.2 Overall Likelihood Ratio Results: Uncalibrated ....................................................... 250
8.5.3 Overall Likelihood Ratio Results: Calibrated ............................................................ 254
8.6 Discussion......................................................................................................................................... 258
8.6.1 Do Clicks Improve the Complete System? .................................................................. 258
8.6.2 Comparing Individual Parameters to the Systems ................................................. 260
8.6.3 Limitations .............................................................................................................................. 263
8.7 Conclusion ........................................................................................................................................ 265
Chapter 9: Discussion .......................................................................................................................... 267
9.1 Summary of the Forensic Speaker Comparison Practices Survey ............................ 267
9.2 Summary of Phonetic/Linguistic and Forensic Findings for Individual Parameters
...................................................................................................................................................................... 268
9.2.1 Articulation Rate ................................................................................................................... 268
9.2.2 Long-Term Formant Distributions ................................................................................ 269
9.2.3 Long-Term Fundamental Frequency............................................................................ 271
9.2.4 Click Rate.................................................................................................................................. 272
8
9.3 Summary of Discrimination Performance by the Overall System ............................ 274
9.4 Overall Findings ............................................................................................................................. 276
9.4.1 Human-Based System versus ASR ................................................................................. 276
9.4.1.1 A Fair Comparison? .................................................................................................... 277
9.4.1.2 Scope for Improvements in the Human-Based System ............................... 279
9.4.1.3 The Trade-Off ............................................................................................................... 280
9.4.2 Obstacles Facing the Implementation of an LR Framework ............................... 281
9.5 Methodological Limitations ...................................................................................................... 283
9.6 Implications for Forensic Speaker Comparisons ............................................................. 285
Chapter 10: Summary & Conclusion............................................................................................ 287
10.1 Research Questions Revisited ............................................................................................... 287
10.2 Opportunities for Future Research ..................................................................................... 291
10.3 Conclusion ..................................................................................................................................... 292
Appendices ................................................................................................................................................ 293
List of Abbreviations ............................................................................................................................ 297
References ................................................................................................................................................. 299
9
List of Tables and Figures
List of Tables
Table 3.1: Methods of analysis employed by country 78

Table 3.2: Places of work against method of analysis employed 78
Table 3.3: Methods used for analysis in forensic speaker comparisons 81
against conclusion frameworks
Table 3.4: Conclusion frameworks used by country 82
Table 3.5: Satisfaction with conclusion framework 82
Table 3.6: Creation of guidelines/protocols 83
Table 3.7: Workers involved in a single case 84
Table 3.8: Frequency of consonant analysis in English 86
Table 4.1: Overview of articulation rate studies 101
Table 4.2: Examples of memory stretches for Speaker 036 108
Table 4.3: Summary of articulation rate statistics when varying the 120
minimum syllables in the speech interval
Table 4.4: Within-speaker differences for articulation rate 121
Table 4.5: Between-speaker differences for articulation rate 122
Table 4.6: Expressions for strength of evidence in terms of Log10 LR and 126
corresponding verbal expression following Champod and
Evetts (2000) verbal scale
Table 4.7: Summary of LR-based discrimination for articulation rate 128
(100 speakers)
Table 4.8: Summary of LR-based discrimination for mean articulation 131
rate when varying the minimum number of syllables in a
speech interval (25 speakers)
Table 5.1: Overall between-speaker results for LTFD1-4 162
Table 5.2: Overall within-speaker results for LTFD1-4 162
Table 5.3: Summary of LR-based discrimination for LTFD1-4 (100 165
speakers)
Table 5.4: Summary of LR-based discrimination for different LTFD1-4 167
10
combinations (100 speakers)
Table 5.5: Package length variability 171
Table 5.6: Overview of discriminant formant studies where F3 performs 173
best
Table 5.7: Summary of LR-based discrimination for LTFD and MFCC in 175
competing studies
Table 6.1: Tailored Frequency Ranges for Selected Speakers 189
Table 6.2: Summary of LR-based discrimination for F0 (100 speakers) 196
Table 6.3: F0 package length variability for LR results 198
Table 6.4: Fundamental frequency across different package lengths 198
Table 7.1: Number of clickers versus number of non-clickers over 213
varying speech sample lengths
Table 7.2: Summary of Speakers Mean and Median Click Rates - Int1 223
versus Int2 and Int3
Table 7.3: Changes in click rate across speaker - Int1 versus Int2 224
Table 7.4: Changes in click rate across speaker - Int1 versus Int3 224
Table 7.5: Mean click rates of the three interlocutors 224
Table 8.1: Formant pairing within LTFD 237
Table 8.2: Between-parameter pairings 237
Table 8.3: Correlation coefficients within LTFD 242
Table 8.4: Correlation coefficients within- and between-parameters 247
Table 8.5: Summary of LR-based discrimination for the complete system 250,
(100 speakers) 288
Table 8.6: Summary of LR-based discrimination for alternative systems 253
(100 speakers)
Table 8.7: Summary of LR-based discrimination for individual 260
parameters (100 speakers)
Table 8.8: Performance comparison between individual parameters and 261
the complete system
Table 8.9: Best-performing system or individual parameter in relation to 262
LR statistics
Table 9.1: Human-based results against ASR (Batvox) results from 275
French et al. (2012) on studio quality data
11
List of Figures
Figure 2.1: Illustration of the UK Position Statement (Rose and Morrison, 55

2009, p. 141)
Figure 4.1: Distribution of mean articulation rates 110
Figure 4.2: Distribution of standard deviations in articulation rate 111
Figure 4.3: Cumulative percentages for mean articulation rate 112
Figure 4.4: Cumulative percentages for standard deviation in articulation 113
rate
Figure 4.5: Comparison of mean articulation rate for memory stretches 117
versus inter-pause stretches
Figure 4.6: Tippett plot of articulation rate 130
Figure 5.1: Distribution of mean LTFD1 147
Figure 5.2: Distribution of standard deviation in LTFD1 148
Figure 5.3: Cumulative percentages for mean LTFD1 149
Figure 5.4: Cumulative percentages for standard deviation of LTFD1 150
Figure 5.8: Cumulative percentages for standard deviation in LTFD2 153
Figure 5.10: Distribution of standard deviation distribution in LTFD3 155
Figure 5.17: Tippett plot of LTFD1+2 168
Figure 5.18: Tippett plot of LTFD2+3 169
Figure 5.19: Tippett plot of LTFD1+2+3 169
Figure 5.20: Tippett plot of LTFD1+2+3+4 170
Figure 6.1: Example of a text grid annotation 188
12
Figure 6.2: Distribution of mean fundamental frequency 190
Figure 6.3: Distribution of standard deviation in fundamental frequency 191
Figure 6.4: Cumulative percentages for mean fundamental frequency 192
Figure 6.5: Cumulative percentages for standard deviation in 193
fundamental frequency
Figure 6.6: Tippett plot of fundamental frequency 197
Figure 7.1: The action of the vocal organs in producing a velaric 204
ingressive voiceless dental click k (a) first stage velic and
anterior closure; (b) second stage, expansion of the enclosed
oral space; (c) third stage, release of the anterior closure.
(Laver, 1994, p. 176)
Figure 7.2: IPA Chart - Clicks Excerpt 204
Figure 7.3: Distribution of click occurrences by functional category 212
Figure 7.4: Distribution of click totals over five minutes of speech 214
Figure 7.5: Distribution of click rate (clicks/minute) in DyViS population 215
Figure 7.6: Cumulative percentages for click rate 216
Figure 7.7: Mean and range of click rates across all speakers 218
Figure 7.8: Distribution of standard deviation in click rate 219
Figure 7.9: Cumulative percentages for standard deviation in click rate 220
Figure 8.1: LTFD1 versus LTFD2 238
Figure 8.7: Mean AR versus LTFD1-4 243
Figure 8.8: Click rate versus mean AR 244
Figure 8.9: Click rate versus mean F0 244
Figure 8.10: Click rate versus LTFD1-4 245
Figure 8.11: Mean F0 versus LTFD1-4 246
Figure 8.12: Mean F0 versus mean AR 247
Figure 8.13: Tippett plot of the complete system 251
Figure 8.14: Zoomed-in Tippett plot of the complete system 252
13
Figure 8.15: Tippett plot of the complete system - parameters calibrated 255
individually and then combined
Figure 8.16: Zoomed-in Tippett plot of the complete system - parameters 256
calibrated individually and then combined (-6 to 6 LLR)
Figure 8.17: Zoomed-in Tippett plot of the complete system - system 257
calibration after combination of parameters (-10 to 10 LLR)
14
List of Appendices
Appendix A: Survey Instructions
Appendix B: Example of a Bayesian Network for Speech Evidence
15
Acknowledgements
There is an ancient African proverb that says it takes a village to raise a child.
A portion of that proverb it takes a village- is now colloquially used to
acknowledge the influence a group of people can have when contributing to
something bigger than themselves.
Despite appearing as the single author of this piece of work, I believe that it
takes not just a single individual but a village to bring a PhD to fruition. To
that extent, I would like to take the opportunity to thank my village.
The first thank you goes to my supervisor, Professor Peter French, whose
support, guidance, and encouragement has helped shape every bit of this
research. Thank you for every discussion (of which there are far too many to
even count), the time you have spent editing drafts/abstracts/papers, and all
the opportunities you have graciously allowed me.
Thank you is also extended to Professor Paul Foulkes and Dr. Dominic Watt for
their advice and suggestions on this thesis. A general thank is also due to the
York team for helping to secure the Universitys position in the BBfor2 project.
I must thank BBfor2 for being a wonderful learning environment over the past
years. An important thank you goes to Dr. David van Leeuwen and Dr. Henk van
den Heuvel for their leadership and organization of the project. A thank you is
also extended to my supervisor in the BBfor2 network, Dr. Didier Meuwly, for
providing valuable insight into the application of Bayes Theorem in forensics.
During my PhD I was fortunate enough to go on two placements, from which I

learned so much. Thanks go to those who helped me at the Netherlands
Forensic Institute, especially Dr. Marjan Sjerps, Jacob de Zoete, and Vikram
Doshi with all things Bayesian.
I would also like to thank those at the University of Canterbury in the New
Zealand Institute for Language, Brain and Behaviour, especially Professor Jen
Hay for allowing me to be a part of such an amazing linguistic environment.
Thank you to Dr. Kevin Watson and Dr. Lynn Clark for giving me the
opportunity to work alongside you both.
A special thank you goes to Professor Colin Aitken, to whom I sent an email
containing a complicated statistics question in early 2012. If I could have
anticipated what was to follow, then I probably would have emailed sooner.
That email began what was to become an extremely rewarding exchange of
ideas at the interface of phonetics and statistics. Thank you for taking the time
to reply.
16
A thank you is also extended to those who have helped me along the way
through discussions, insight, and simply encouragement. Thank you to
Professor Anders Eriksson, Phil Harrison, Dr. Michael Jessen, Dr. Christin
Kirchhbel, Dr. Carmen Llamas, Professor Francis Nolan, Dr. Richard Ogden (for
all things non-pulmonic), and Lisa Roberts.
I must also express my gratitude to all 36 anonymous forensic experts who

participated in my survey for this PhD research. Without them my PhD would
lack a solid basis.
To all my office buddies over the years, I thank you for putting up with my
annoying finger tapping (as I counted syllables for AR) and the clicking sounds I
made (trying to determine place of articulation), the discussions we have had,
and simply from keeping me from talking to myself: Natalie Fecher, Jessica
Wormald, Becky Taylor, and Rana Alhussein Almbark.
A thank you must also go to Vincent Hughes for all our LR discussions. Without
a fellow LR-researcher this thesis just would not be the same. A thank you also
goes to the Forensic Research Group in the Language and Linguistic Science
Department at the University of York for providing a safe and nurturing
environment in which to exchange ideas.
Now to those behind the scenes. I must thank my family - Mom, Dad, and Lauren
for your unequivocal love and support. You have always encouraged me to
follow my interests, so much so that your encouragement has never faltered
even with me being 6,000 miles away from home. And sorry for all those middle
of the night phone calls when I miscalculated the time difference.
To Tom, my best friend and biggest supporter, you know the details of this PhD
almost as well as I do. Thank you for being my shoulder to cry on when things
got stressful, lending your ear as I discussed a long days worth of work,
providing unyielding patience, and your continued support and unconditional
love.
Finally, despite my love for phonetics and linguistics, the work reported in this
thesis would not have been possible without the financial support of the
Bayesian Biometrics for Forensics (BBfor2) Network, funded by Marie Curie
Actions (EC Grant Agreement No. PITN-GA-2009-238803). Without the
generous support of this funding body, this thesis would not be possible.
To my village I thank you.
17
18
Declaration
This is to certify that this thesis comprises original work and that all
contributions from external sources are acknowledged by explicit references.
I also declare that aspects of the research have been previously published or
submitted to journals and conference proceedings. These publications are as
follows:
o Gold, E. and Hughes, V. (2014). Issues and opportunities for the

application of the numerical likelihood ratio framework to forensic
speaker comparison. Science and Justice. <http://www.sciencedirect.
com/science/article/pii/S1355030614000501>
o Hughes, V., Brereton, A., and Gold, E. (2013). Sample size and the
computation of numerical likelihood ratios using articulation rate. York
Papers in Linguistics, 13, pp. 22-46.
o Gold, E., French, P., and Harrison, P. (2013). Clicking behavior as a
speaker discriminant in English. Journal of the International Phonetic
Association, 43(3), pp. 339-349.
o Aitken, C.G.G. and Gold, E. (2013). Evidence evaluation for discrete
data. Forensic Science International, 230(1-3), pp. 147-155.
o Aitken, C.G.G. and Gold, E. (2013). Evidence evaluation for multivariate
discrete data. In Proceedings of the 59th ISI World Statistics Congress,
Hong Kong. < http://www.statistics.gov.hk/wsc/IPS021-P1-S.pdf>
o Gold, E., French, P., and Harrison, P. (2013). Examining long-term
formant distributions as a discriminant in forensic speaker comparisons
under a likelihood ratio framework. In Proceedings of Meetings on
Acoustics, (POMA - ICA 2013, Montreal) 19. [DOI: 10.1121/1.4800285]
o Gold, E. & French, P. (2011). International practices in forensic speaker
comparison. International Journal of Speech, Language and the
Law, 18(2), pp. 293-307.
o Gold, E. & French, P. (2011). An international investigation of forensic
speaker comparison practices. In Proceedings of the 17th International
Congress of Phonetic Sciences, Hong Kong, pp. 751-754.
Signed:
Erica Gold
Date: January 13, 2014
19
20
Not everything that can be counted counts
and not everything that counts can be counted.
-William Bruce Cameron
21
22
Chapter 1 Introduction
The research presented in this thesis explores the calculation of numerical
likelihood ratios using both phonetic and linguistic features. Articulation rate,
fundamental frequency, long-term formant distributions, and clicks (velaric
ingressive plosive sounds) are analyzed with the purpose of considering intra-
and inter-speaker variation, levels of speaker discrimination, the strength of
evidence, and the viability of presenting forensic speaker comparison
conclusions as numerical likelihood ratios. This chapter outlines the
contribution of the thesis in the field of forensic speech science, and provides a
short summary of forensic speaker comparison and conclusion frameworks
used in forensic speaker comparison cases. The research aims are then
described, and overviews of each chapter are provided.
1.1 Forensic Speaker Comparison
Forensic speaker comparison (FSC) is noted as being the most common
task carried out by forensic phoneticians (Foulkes and French, 2012, p. 558),
and the majority of research in the field of forensic speech science (hereafter
FSS) is oriented towards this task. FSC is also referred to by other terms such as
(forensic) speaker identification, (forensic) speaker recognition, and (forensic)
voice comparison (Rose, 2002; Rose and Morrison, 2009). However, the term
comparison is preferred in this thesis for two reasons. Firstly, it is not possible
to achieve an identification with 100% certainty under a frequentist 1
1The Merriam-Webster dictionary defines frequentist as [defining] the probability of an event

(as heads in flipping a coin) as the limiting value of its frequency in a large number of trials
<http://www.merriam-webster.com/dictionary/frequentist [Accessed 10 January 2014].
23
conclusion framework, given that there is always, to some extent, within-
speaker variability. Secondly, under a Bayesian framework the expert should
not take on the role of trier of fact by providing an identification. Rather,
his/her responsibility is to express the probability of obtaining the evidence
under the hypothesis that the samples came from the same person, versus the
probability of obtaining the evidence under the hypothesis that two different
speakers produced the criminal and suspect samples. The term speaker is
preferred over voice in this thesis as not all parameters examined in FSC work
are products of just the voice per se. The manifestation of speech parameters
can also be a reflection of the social and psychological mind-set of the individual
(e.g. French et al., 2010). Therefore, forensic speaker comparison is the term
preferred over other possible naming conventions.
The analysis in an FSC typically involves the comparison of two (or
more) recordings: a criminal sample (also referred to as an unknown,
disputed, trace, or questioned sample) and a suspect sample (also referred to
as a known or reference sample). The criminal sample is a recording adduced
as evidence that contains the speech of an unknown individual. It is possible for
the criminal recording to also contain other sounds associated with the crime
taking place. In the UK, the suspect sample is usually a recording of a police
interview (Nolan, 1983; Rose, 2002) with the suspect. The objective of the
expert forensic phonetician is to provide the trier(s) of fact with an informed
opinion regarding the probability of obtaining the evidence (the
similarities/differences between the criminal and suspect samples) under the
hypothesis that the samples came from the same person, versus the probability
of obtaining the evidence (the typicality of the analyzed speech parameters)
24
under the hypothesis that two different speakers produced the criminal and
suspect samples. This objective can be reached by experts using a variety of
methods (e.g. acoustic analysis, auditory analysis, acoustic and auditory
analysis, fully automatic speaker recognition2 (ASRs), or human-assisted ASR);
however, the most common method employed by experts is the combination of
auditory phonetic and auditory acoustic analysis of phonetic, linguistic, and
non-linguistic speech parameters (e.g. laughter, coughs; French et al., 2010).
1.1.1 Expression of Conclusions
Just as there is variability in the methodologies preferred for the
comparison of speakers, variability also exists across analyses with regard to
the expression of a conclusion at the end of a FSC. Conclusion frameworks can
include binary decisions (either the two speakers are the same person or they
are different speakers), classical probability scales (probability of identity
between the criminal and suspect; Broeders, 1999), the UK Position Statement
(a potentially two-part decision based on assessing consistency and
distinctiveness of the samples; French and Harrison, 2007), and likelihood
ratios (LR, either verbal or numerical, expressing the likelihood of finding the
evidence given a same-speaker versus different-speaker hypothesis; Morrison,
2009b; 2009c). There has recently, however, been a strong promotion of the
use of the LR framework, as it is advanced as being the only logically and
legally correct framework (Rose and Morrison, 2009, p. 143).
2Other researchers have use ASR to mean automatic speech recognition (e.g. Goel, 2000).
However, ASR is used in this thesis to mean automatic speaker recognition.
25
The use of the numerical LR in research literature has been given
increasing attention, starting with Rose (1999). However, the use of the
numerical LR in courts in cases involving FSCs is rare (see Rose, 2012; 2013).
Many reasons exist for the limited representation of the numerical LR in court, a
number of which are addressed in French et al. (2010). However, the key
practical limitation cited by French et al. (2010) preventing the implementation
of numerical LR conclusions lies in the limited availability of population
statistics and the difficulty of collecting them. If numerical LRs were to be
calculated with regard to the (very) few parameters for which there are
available population statistics, French et al. (2010, p. 149) argue that one runs
the risk of producing an opinion that could lead to a miscarriage of justice. This
is due to the fact that the analysis would fail to consider a large number of other
available parameters for which there are no population statistics. In turn, this
could impact the conclusion the expert would arrive at in a FSC case.
The motivation for this thesis stems directly from the difficulties and
limitations associated with calculating a numerical LR and the increased desire
for the field of forensic speech science to align itself with other more developed
disciplines of forensic science (e.g. DNA). Previous discussions have tended to
focus simply upon the reasons for or against the implementation of a numerical
LR. However, far less empirical work has been carried out to examine the
practicalities associated with the calculation of numerical LRs. The present
study serves as an exercise in calculating numerical LRs for speech data. The
data are derived from the speech of a homogeneous group of speakers, while
assessing the discriminant ability of these parameters in combination. This
exercise is intended to parallel the methodologies and procedures that would
26
be implemented in a real FSC case, should the analyst choose to utilize a
numerical LR framework.
1.2 Research Aims
The primary aims of this thesis are threefold. The first aim is to provide
the field of FSS with a comprehensive summary of current FSC practices used
around the world, which has never been previously available. The survey will
provide details regarding the types of analysis that are used and their
frequency, information on the speech parameters employed, experts opinions
of the discriminant value of those parameters, and the conclusion frameworks
they adopt. The survey will also serve as the primary motivation for the
selection of the four phonetic and linguistic parameters examined in this thesis.
The second aim is to expand the breadth of FSS literature by examining
the actual discriminant value of parameters (individually and in combination)
identified by expert forensic phoneticians as being good speaker discriminants.
Irrespective of the expectations and actual level of discrimination potential
carried by these given parameters, their distributions within the analyzed
population will nonetheless provide the field of forensic phonetics with useful
information that can further inform FSC casework. The analysis focuses on
three intrinsically quantitative phonetic parameters and one ostensibly
qualitative linguistic parameter, while aiming to increase the number of speech
parameters that can be considered under a numerical LR. Through this analysis,
the thesis simultaneously aims to contribute detailed population statistics for
four phonetic and linguistic parameters in a large, homogeneous group of
27
speakers. As such, this research will address the arguments set forth in French
et al. (2010) about the limited availability of population statistics.
The third, and final, aim of the present body of work is to take the
necessary steps to assess the practical limitations and opportunities associated
with the implementation of a numerical LR framework in FSCs. This begins by
examining potential correlations that exist between and within parameters, and
is then followed by appropriately combining the individual pieces of speech
evidence. Numerical LRs are then calculated, and strength of evidence and the
performance of the combined system are considered. Potential pitfalls and
successes are then acknowledged, as doing so is necessary in contributing to the
on-going discussion of whether it is practical to adopt a numerical LR
framework (in part or full) for FSCs. Most importantly, this work aims to
provide a transparent and objective assessment of the viability of implementing
a numerical LR framework for FSCs. This assessment will be approached from a
structured learning (data-driven) perspective, rather than through giving
subjective theoretical opinions regarding the use of numerical LRs in FSC
casework.
1.3 Thesis Outline
In Chapter 2, an overview is provided of relevant literature relating to the so-
called paradigm shift in forensic science, changes in the law Bayes Theorem
FSCs, the use of likelihood ratios (LRs) in FSCs, and a discussion of the
limitations relating to these topics. The research questions for this thesis that
have arisen from previous research reviewed in this chapter are also presented.
28
Chapter 3 reports on the construction, contents, and results of the first
comprehensive international survey of forensic speaker comparison practices.
It provides a summary of current practices around the world, commonly-used
phonetic, linguistic, and non-linguistic parameters in casework, conclusion
frameworks, and expert opinion about which parameters are believed to be
highly discriminant. A selection of those parameters found or claimed to be
useful speaker discriminants by survey participants is chosen for further
examination in subsequent chapters. They include: articulation rate, long-term
formant distributions, fundamental frequency, and clicks (velaric ingressive
plosives).
The investigation of articulation rate (AR) as a speaker discriminant is
presented in Chapter 4. A summary of the AR literature is provided, followed by
an analysis of the distribution of AR in the test population. This chapter
explores the methodologies used for calculating AR by manipulating the
minimum syllable length requirement in a speech interval and comparing inter-
pause stretches with memory stretch intervals. The chapter concludes by
calculating LRs for AR and determining the levels of discrimination, strength of
evidence, and validity of the system.
Chapter 5 analyzes long-term formant distributions (LTFD) as a speaker
discriminant. A summary is provided of research that explores LTFD as a
parameter for FSC. LTFDs are analyzed individually as well as in combinations
relevant to forensic casework, while also providing population statistics. The
chapter investigates the effects that the package length (time intervals) of
tokens may have on results. Finally, LRs are calculated for individual formants
29
as well as in combination. The levels of discrimination are presented, along with
strength of evidence and validity of the system.
Long-term fundamental frequency (F0) as a speaker discriminant is
explored in Chapter 6. The chapter offers a summary of the relevant literature
on F0 as well as the external factors known to affect F0. Population statistics are
offered for F0, and the effects of the package length of tokens are investigated
for potential differences in the discriminant results for F0. The chapter
concludes with the calculation of LRs, and examines the levels of discrimination,
strength of evidence, and validity of F0 as a system.
Chapter 7 analyzes clicks3 (the final parameter in the thesis) as a speaker
discriminant. A summary of the literature on clicks in general is provided at the
beginning of the chapter, as well as clicks in conversation analysis. This chapter
considers click rate (frequency of velaric ingressive plosives) as a discriminant
parameter, and provides population statistics for within- and between-speaker
variability. The effects of accommodation are explored in relation to increases
in within speaker variation. The chapter concludes by discussing the impeding
limitation of not being able to calculate LRs for click rate, due to the lack of
appropriate modeling techniques for the data distribution presented by click
rate.
The correlations and combinations of those parameters presented in
Chapters 4-7 are explored in Chapter 8. The chapter provides a summary of the
literature on correlations and combinations of parameters in FSC. Correlations
are calculated between all parameters as well as within parameters. These data
are used to inform the appropriate methods for the combination of parameters
3 Used as discourse markers in conversation.
30
in order to create a complete system (consisting of the four parameters
explored in this thesis). Overall likelihood ratios (OLRs) are calculated for the
complete system as well as ten alternative systems that consist of different
combinations of LTFD, F0, and AR. The performance of the complete system (in
terms of strength of evidence and validity) is discussed in comparison to the
performance of the alternative systems.
The results presented in Chapters 3-8 are considered collectively and
discussed in Chapter 9. A comparative analysis of individual parameters is
offered alongside the combination of parameters as a system, examining levels
of discrimination between speakers, strength of evidence, and validity. The
phonetic-linguistic (human-based) system consisting of AR, LTFD, F0, and clicks
is then compared to the performance of ASRs. To conclude, limitations
associated with the calculation of numerical LRs are discussed, as well as the
implications for using a numerical LR framework in casework.
Finally, Chapter 10 provides a summary of the overall findings of the
thesis, revisits the thesis aims and identifies opportunities and challenges that
face the implementation of a numerical LR should practitioners choose to adopt
such a FSC conclusion framework.
31
Chapter 2 Literature Review
In this chapter, an overview is presented of the literature surrounding the so-
called paradigm shift in forensics changes in the law Bayes Theorem forensic
speaker comparisons (FSCs), the use of likelihood ratios (LRs) in FSCs, and the
limitations and shortcomings surrounding these topics that have led to the
research questions of this thesis. All subsequent chapters contain a literature
review concerning the issue or parameter under focus.
2.1 The Paradigm Shift
The term paradigm shift was first introduced by Kuhn in 1962. A paradigm in
the sciences is defined by Kuhn as a conceptual framework that only members
of a particular scientific community share. He goes on to describe a paradigm
shift as a change in these basic assumptions within the ruling theory of science.
An example of one of the most famous paradigm shifts in science is the
transition from a Ptolemaic cosmology to a Copernican one, in which the sun is
the center of the universe rather than the Earth (Kuhn, 1962). Kuhn (1962)
argues that once a paradigm shift is complete, a scientist is unable to reject the
new paradigm in favor of the old one. As asserted by Kuhn (1962), paradigms
exist in all (sub)domains of science, and forensic science is no exception.
In 2005, Saks and Koehler wrote a review entitled The Coming Paradigm
Shift in Forensic Identification Science in which they argued that traditional
forensic sciences should replace antiquated assumptions of uniqueness and
perfection with a more defensible empirical and probabilistic foundation (Saks
and Koehler, 2005, p. 895). They begin their review by describing the state of
32
traditional forensic science that follows a frequentist view of evaluation,
whereby a decision is made on the probability of a single hypothesis (and
without considering prior probabilities). Typically that hypothesis would be
that two evidentially-relevant traces, e.g. recorded speech samples, were made
by a single object/person (Osterburg, 1969; Stoney, 1991). This form of a
hypothesis links evidence to a single object or person to the exclusion of all
others in the world (Saks and Koehler, 2005, p. 892) Linking evidence to a
single object or person is done based on the assumption of uniqueness, whereby
the idea is that two evidentially-relevant traces produced by different people or
objects will always be different. Therefore when two pieces of evidence are
being compared that are not observably different, an expert will conclude that
they were made by the same object or person (Saks and Koehler, 2005). The
authors are implicitly drawing attention to the single-hypothesis, frequentist-
paradigm in which the typicality of a piece of evidences characteristics (in a
given population) has failed to be taken into account (i.e. they reject the
uniqueness assumption).
Saks and Koehler (2005) reveal that in the decade leading up to their
review, many people had been falsely convicted of serious crimes, only to be
later exonerated by DNA evidence that had not been previously tested at the
time of the trial. The authors state that erroneous convictions sometimes occur,
and surprisingly in an analysis of 86 cases (ones which resulted in false
convictions), it was found that 63% were due in some part to erroneous forensic
science expert testimony (Saks and Koehler, 2005, p. 892). This was the second
biggest contributing factor to false convictions, after misleading eyewitness
33
identifications4. The authors explicitly state that the criticism does not apply to
DNA evaluation as it is currently practiced; rather, DNA should serve as a model
for other forensic science disciplines. The reasoning behind the statement is
that DNA typing follows three main principles, (1) the technology [is] an
application of knowledge derived from core scientific principles (2) the courts
and scientists [can] scrutinize applications of the technology and (3) it offers
data-based probabilistic assessments of the meaning of evidentiary matches
(Saks and Koehler, 2005, p. 893).
The authors strongly advise practitioners of other forensic disciplines to
emulate the approach taken by DNA typing, whereby the courts are provided
with quantifiable evidence, error rates of the technology, and match
probabilities being calculated from two competing hypotheses. Without
explicitly stating it, Saks and Koehler are essentially arguing for the adoption of
the likelihood ratio (LR; 1.1.1) as the medium for presenting conclusions to
the trier(s) of fact (e.g. judge, jury). They are arguing for forensic science to
move from a pre-science to an empirically grounded one that will be
transparent and properly scientific. In order for other forensic disciplines to
take on such an approach, Saks and Koehler (2005, p. 892) recommend that
forensic scientists will need to work closely with experts in other fields to
develop efficient methods.
In recent years, following the paper by Saks and Koehler (2005), there
have also been calls for improvements in the quality of forensic evidence by a
number of legal and government bodies. It has been argued that all areas of
4Multiple factors were considered in each false conviction. Therefore, while erroneous forensic
evidence was a contributing factor in 63% of the false convictions, erroneous eyewitness
testimony was the only other factor contributing to more false convictions (71%; Saks and
Kohler, 2005, p. 892).
34
forensic science need to be more transparent, that forensic examinations should
be based on validated methodologies, and that the results should be replicable
and expressed in quantitative terms (U.S. National Research Council, 2009;
House of Commons Northern Ireland Affairs Committee 2009; Law
Commission of England & Wales, 2011). These calls for changes to forensic
evaluation were made for the same reasons that Saks and Koehler (2005)
alluded to with respect to false convictions being made from poorly presented
forensic evidence as well as the changes that have occurred in the law.
2.2 Changes in the Law
A number of rulings made in the last century have significantly changed the face
of expert evidence evaluation and testimony in various countries, especially the
United States. Starting in 1923, with the ruling of Frye v. United States, courts
moved away from accepting testimony from expert witnesses on the basis of the
experts academic pedigree. Rather, a change was made by a federal appellate
court, which rendered expert evidence inadmissible when it was based on
methods not used by others in the same forensic discipline. The Frye ruling
(Frye v United States (293 F. 1013 D.C. Cir. [1923])) determined that expert
testimony was only admissible if the method of analysis used gained general
acceptance in the particular field in which it belongs. (Frye v United States,
paragraph 5).
In 1993, the law changed once again as scientific methods continued to
improve. In Daubert v. Merrell Dow Pharmaceuticals (509 US 579 [1993]), the
United States Supreme Court implemented a new ruling with regard to the
admissibility of forensic evidence, whereby the forensic science in question
35
must demonstrate that it can stand on a dependable (i.e. tested) foundation. The
ruling challenged those in the field of forensics to show that the forensic
method in question had been tested, that its error rate has been established,
and this error rate was acceptably low. The Daubert ruling has since been
interpreted to mean that forensic sciences should be quantifiable, validated, and
reliable. The ruling by the United States Supreme Court was intended to lower
the threshold of admissibility for new and cutting-edge methodologies, which
would have previously been considered inadmissible under the Frye ruling. At
the same time, Daubert was meant to raise the threshold for long-established
methods lacking a proper scientific foundation. Daubert subjected forensic
sciences to serious methodological scrutiny for the first time.
By 1995, in the case of the United States v. Starzecpyzel, a loophole in the
Daubert ruling was brought to light. The case in question included handwriting
identification expertise, where a federal district court concluded that
handwriting identification had no scientific basis, following the Daubert ruling.
This decision was made even though the field of handwriting analysis had
dedicated certification programs and professional journals. However, due to the
loophole in Daubert the handwriting evidence was not excluded. The reason
given was that since the methods used to collect evidence were found to have no
scientific basis, Daubert did not apply to handwriting identification as it was not
viewed as scientific evidence. The case of Starzecpyzel gave precedent for
providers of other forensic testimony to find a way around Daubert by lowering
the threshold for admissibility and declaring weakly-founded forensic
testimony as non-scientific, thus bypassing the Daubert ruling altogether.
36
It was not until 1999, in the case of Kumho Tire v. Carmichael, that the
United States Supreme Court directly addressed whether or not Daubert applied
to non-sciences. A brief was put together by a number of law enforcement
organizations in which they argued that the majority of the expert testimony
that they offered did not include scientific theories, methodologies, techniques,
or data (Brief Amicus Curiae of Americans for Effective Law Enforcement,
1997). This was stated in relation to the testimony of specific fields of
investigation, such as: accident reconstruction, fingerprint, footprint and
handprint [identification], handwriting analysis, firearms markings and
toolmarks, bullets, and shell casings, and bloodstain pattern identification (Brief
Amicus Curiae of Americans for Effective Law Enforcement et al., 1997).
Ironically, the practitioners that were initially lobbying for their expertise to be
admissible on scientific grounds were now denying that they were a science.
Despite efforts to maintain the non-science loophole of Daubert the United
States Supreme Court ruled in Kumho Tire that all expert testimony would be
required to pass appropriate tests of validity (set forth by Daubert) in order to
be admissible in court.
Although the rulings described in this section pertain to law in the United
States, these rulings have had a large impact on the legislation in other
countries. In the United Kingdom, expert testimony is typically admissible on
the basis of the qualifications of the expert testifying rather than the methods
employed by that expert. This principle was influenced by the case of R v.
Bonython in Australia where the Supreme Court ruled on the admissibility of
handwriting evidence. They concluded that forensic evidence testimony is
admissible when (i) a layperson is unable to form a sound judgment on the
37
matter without the assistance of the witness possessing special knowledge or
experience in the area and when (ii) the subject matter of the opinion forms
part of a body of knowledge or experience which is sufficiently organised or
recognised to be accepted as a reliable body of knowledge or experience (R v.
Bonython, 1984, paragraph 5). This ruling can be interpreted as being
intermediate between the Frye and Daubert rulings, where the Bonython ruling
encompasses the Frye ruling and includes the expectation that the testimony is
reliable (again, this is usually satisfied with reference to the experts academic
pedigree and the lack of previous miscarriages of justice in relation to the given
expert testimony).
In recent years the House of Commons Northern Ireland Affairs
Committee (2009) and the Law Commission of England & Wales (2011) have
also been influenced by such U.S. rulings, and have urged forensic sciences to
make changes to their current practices that would align them more closely
with measures set out in Daubert. With respect to the changes in legislature, the
developments in presenting DNA typing, the paradigm shift, and the calls made
by legal and government bodies, all of these factors (explicitly or inexplicitly)
are convergent in their desire for forensic disciplines to adopt a Bayesian
framework and to implement likelihood ratios.
2.3 Bayes Theorem
Bayes Theorem was first proposed by Sir Thomas Bayes in the 1740s, then
updated and published by Richard Price (Bayes and Price, 1763), and later the
same principles were rediscovered and updated further by Pierre Simon
38
Laplace5 (Laplace, 1781). Laplace is in reality the man who turned Bayes
Theorem into the modern-day scientific application that is currently used
around the world (Bertsch McGrayne 2012). Bayes Theorem was created as a
way in which to update ones beliefs. The theorem has three central
components: the posterior odds, the prior odds, and the likelihood ratio (LR), as
illustrated in Equation 1. The components of Equation 1 are explained in detail
in the subsequent sections with respect to forensic science.
(1)
( | )
Posterior Odds Prior Odds Likelihood

Ratio
Adapted from: Aitken and Taroni (2004, p. 95)
In Equation (1), p represents the probability, where Hp is the prosecution
hypothesis (e.g. the criminal and suspect are the same person) and Hd is the
defense hypothesis (e.g. the criminal and suspect are different people). The E in
Equation (1) is representative of the evidence in question. Bayes Theorem
proposes that the posterior odds (the probability of the prosecution hypothesis
5The term updated is used here to mean that a prior probability can be adjusted/modified by
taking into account any new evidence or observation (e.g. likelihood ratio(s) in forensics) to
arrive at a posterior probability (see Equation 1).
39
being correct given the evidence divided by the probability of the defense
hypothesis being correct, given the evidence) is equal to the prior odds (the
probability of the prosecution hypothesis being correct divided by the
probability of the defense hypothesis being correct; see 2.3.2 for an example)
multiplied by the likelihood ratio (the probability of obtaining the evidence
given the prosecution hypothesis divided by the probability of obtaining the
evidence given the defense hypothesis; see 2.3.1 for an example).
2.3.1 Likelihood Ratio
The likelihood ratio (LR) is a gradient measure of the value of evidence
(Aitken and Taroni, 2004) or what is also referred to as the strength of evidence
(Rose, 2002) under a Bayesian framework. An LR is the calculation of the
probability of obtaining the results of a given forensic examination on the basis
of the prosecution hypothesis divided by the probability of obtaining those
same results on the basis of the defense hypothesis. The LR is the only portion
of the Bayesian framework in which a forensic expert should provide an
opinion. The opinion that the expert provides on the strength of evidence is
calculated from two competing probabilities. It should be noted that calculating
an LR does not constitute a Bayesian exercise in and of itself (i.e. it only
constitutes one part of the Bayesian framework), as that would imply the
additional consideration of prior odds (Champod and Meuwly, 2000).
The numerator of the LR is the probability of the prosecutors
hypothesis (the evidence being from the same person/object), while the
denominator is typically the probability of the defense hypothesis (the evidence
has come from a different person/object). Ideally, the defense hypothesis would
40
be set by the defense (Champod and Meuwly, 2000); however, this is rarely
done and the responsibility usually falls to the expert, who typically renders the
hypothesis as the evidence came from someone/something else in the world.
The defense hypothesis has also been referred to as a random match
probability (Champod and Meuwly 2000, p. 195). In the LR equation, when the
numerator presents a greater value than the denominator, there is support for
the prosecution hypothesis, and when the denominator is greater than the
numerator, there is support for the defense hypothesis.
The strength or value of evidence in a case is dictated by the magnitude
of the resulting LR, or rather the distance of the resulting LR from 1. Therefore,
an LR of 100 means that the probability of the evidence (given the competing
prosecution and defense hypotheses) is 100 times more likely to have been
obtained/to have come from the suspect than someone else in the population. If
in the same case the LR was 1/100, then the evidence is 100 times more likely
to have been obtained/come from someone in the population other than the
suspect (Robertson and Vignaux, 1995). The probabilities of the prosecution or
defense hypothesis can take a value between 0 and 1 (inclusive), while the LR
can take a value between 0 and (Aitken and Taroni, 2004). Due to the fact that
LRs can be extremely small (approaching 0) or extremely large (tending
towards ) the LR is often converted into a logarithmic scale with a verbal
translation, which makes it easier for the trier(s) of fact (e.g. judge, jury) to
interpret (Evett, 1995). If an LR is converted using Log10, a positive value then
indicates support for the prosecution hypothesis, while a negative value
indicates support for the defense hypothesis.
41
The combined LRs from multiple pieces of evidence have been referred
to as an overall LR (OLR; Alderman, 2004). If multiple pieces of evidence are
evaluated in a case, individual LRs can be multiplied together (or added
together in the case of Log10 LRs) in order to continue updating an existing
probability following Nave Bayes (Kononenko, 1990; Hand and Yu, 2001).
Nave Bayes refers to when there is an assumption of mutual independence
between the pieces of evidence being combined. When cases of correlated
evidence (predicted through theory or shown through empirical research) are
present, the strength of evidence (resulting LRs) cannot be combined through
simple multiplication, and other methods need to be employed. There are at
present three general remedies for the problem of combining correlated
evidence, (1) the use of an LR algorithm that can handle correlation through
statistical weightings (e.g. the Multivariate Kernel Density LR algorithm; Aitken
and Taroni, 2004), (2) Bayesian networking that will account for correlations by
considering feature distributions and variances and perform statistical
weightings (Aitken and Taroni, 2004), or (3) a solution proposed in the field of
automatic speaker recognition referred to as logistic-regression fusion, which
accounts for correlations in resulting LRs and then applies statistical weightings
(Brmmer et al., 2007; Gonzalez-Rodriguez et al., 2007; Ramos Castro, 2007).
In forensics, an LR can be presented either numerically or verbally. An
example of a numerical LR statement is it is 100 times more probable to obtain
the evidence given the prosecution hypothesis than it is to obtain the evidence
given the defense hypothesis. A verbal LR will not include any numbers in its
statement; instead, different phrases are used to express the strength of
evidence. For example it is more probable to that one would obtain the
42
evidence given the prosecution hypothesis than it would be to obtain the
evidence given the defense hypothesis.
2.3.2 Prior Odds
In Equation 1, the prior odds represent any existing probability of the
hypothesis being true prior to the consideration of new evidence being
introduced; they are then updated by the new information, which results in a
posterior probability (Aitken and Taroni, 2004). For example, suppose a crime
takes place on an island that is inhabited by 101 people. If the perpetrator is
known to be one of the 101 inhabiting the island, then the prior odds of the
suspect being the criminal is 1/100. 6 These prior odds will then be updated by
the trier(s) of fact as new evidence is presented throughout the case. The prior
odds are a key factor in the separation between a frequentist way (see 1.1) of
approaching a problem and a Bayesian way of approaching a problem. A
Bayesian approach allows the probability of a hypothesis to be updated by any
prior probabilities which might affect the posterior probability.
When the prior odds are used in research a numerical value is typically
given, and to incorporate those odds into a Bayesian framework it only requires
simple multiplication with the likelihood ratio. In practice, the prior odds can be
problematic. Robertson and Vignaux (1995, p. 19) state that this is especially
true since very large or very small prior odds can give some very startling
effects. For example, if there were prior odds in a case of , multiplied by an LR
of 4, the posterior odds would be 2 (i.e. in favor of the prosecution hypothesis).
If that same case had prior odds of 1/1000, multiplied by the same LR of 4 (e.g. a
6 Prior odds = p(Hp)/p(Hd), which in this case is p(1/101)/p(100/101) = 1/100
43
Hp of 8 divided by a Hd of 2 is equal to 4), the posterior odds would then be
0.004 (i.e. in favor of the defense hypothesis). Significant (or even small)
changes of the prior odds can dramatically change the posterior odds. This is
demonstrated in the case above, where large prior odds cause the posterior
odds to be in favor of the prosecution hypothesis, and much smaller prior odds
yield posterior odds in favor of the defense hypothesis (despite the LR
remaining constant). Prior odds can also be problematic in practice, given that
Bayes Theorem assigns the responsibility of establishing the prior odds to the
trier(s) of fact. This means that typically the jurors are held to be responsible for
assigning and understanding priors (Robertson and Vignaux, 1995), a task
which is in no way simple.
2.3.3 Posterior Odds
In Equation 1, the posterior odds are the results of the prior odds after
being updated by the LR. Numerically, the posterior odds are the multiplication
of the prior odds by the LR (Aitken and Taroni, 2004). Deriving the posterior
odds, as with the prior odds, is the responsibility of the trier(s) of fact
(Robertson and Vignaux, 1995), and it is up to the trier(s) of fact to determine
the posterior odds by considering the prior odds they had initially established,
in combination with the evidence provided by expert testimony (the LR(s)).
Neither the likelihood ratio nor the prior odds on their own constitute a
Bayesian probability; rather, it is the value of the posterior odds that equates to
a Bayesian belief of probability.
44
2.3.4 Logical Fallacies
There are three main fallacies that can be committed in the
implementation of the Bayesian framework. The first is for the forensic expert
to report posterior odds, since the expert does not typically have access to the
prior odds, as they are generally set by the trier(s) of fact and not an expert
(Rose and Morrison, 2009). Even if the expert were to have access to them, the
prior odds will vary in accordance with individuals personal beliefs about the
cases, and the beliefs are subject to natural bias. This fallacy of presenting
posterior odds, as committed by a forensic expert, also means that the expert is
taking on the role of trier of fact (Robertson and Vignaux, 1995), which in fact
infringes on what has been called the ultimate issue. The ultimate issue in law
is the decision about the guilt or innocence of a suspect by the trier(s) of fact. If
an expert is to present posterior odds, such as an incriminating statement like
the suspect made the shoe mark, the expert then places himself/herself in the
role of decision maker, rather than an objective party presenting facts relating
to the case (see Joseph Crosfield & Sons v. Techno-Chemical Laboratories Ltd.).
The second fallacy is known as the prosecutors fallacy (Thompson and
Schumann, 1987), also referred to as transposing the conditional (Evett, 1995;
Lucy 2005) or the inversion fallacy (Kaye 1993). The prosecutors fallacy
occurs when the probability of the evidence given the hypothesis is
interchanged with the probability of the hypothesis given the evidence (Lucy,
2005). The inversion of the probabilities gives undue weight to the prosecution
hypothesis by assuming that the prior odds of a random match (or two pieces of
evidence found to be similar) are equal to the probability of the defense
45
hypothesis. For example, an expert states that there is a 10% chance the suspect
would have the crime blood type if he were innocent. Thus there is a 90%
chance that he is guilty (Aitken and Taroni 2004, p. 37).
The third fallacy is known as the defenders fallacy (Thompson &
Schumann, 1987), which occurs when minimal weight is attributed to the
evidence. This is done by considering the background population statistics for a
piece of evidence without attention to any associated value (e.g. prior odds). For
example, a DNA profile has a probability of 1% in a total population of 10,000,
which the suspect comes from. The defense argues that the DNA profile would
occur in 100 of these individuals in the population of 10,000, and is therefore of
very little value. On the contrary, cutting the total population from 1 in 10,000
to 1 in 100 means that 9,900 people are being excluded, and also it is highly
unlikely that all of the 100 individuals are equally likely to be the criminal (Evett
and Weir, 1998, p. 32; Lucy, 2005, p. 157).
The three fallacies presented in this section are all flawed in a logical
sense as the expert takes on responsibility that is not his/hers (e.g. presenting
posterior odds), or the prosecutor/defender only considers a portion of Bayes
Theorem in order to arrive at posterior odds on behalf of the trier(s) of fact.
Despite these fallacies having a detrimental effect on the trier(s) of facts
comprehension of the evidence/case with which they have been presented,
these fallacies nevertheless need to be monitored in case miscarriages of justice
occur. For other errors in the interpretation of Bayes Theorem, see Aitken and
Taroni (2004, pp. 78-95).
46
2.4 Forensic Speaker Comparison
Given the current paradigm shift and the logically and legally correct framework
that Bayes Theorem offers, practitioners in the field of FSC are making the
effort to align themselves with other more conceptually advanced areas of
forensics (i.e. those using a Bayesian framework, for example, DNA). Acceptance
of the forensic paradigm shift in FSC has already been acknowledged and
embraced (French et al., 2010). However, the ease with which an LR approach
can be adopted is an issue in itself. This is largely due to the challenges that
speech data in the forensic context present. This section builds upon the
forensic speaker comparison (FSC) introduction in 1.1. It provides further
background information on the complexity of speech data used in FSCs, speech
parameters that are commonly analyzed in FSCs, and the way in which FSC
conclusions are currently framed in the UK. This section will situate the
challenges facing FSC in comparison to other forensic disciplines, while
demonstrating the current state of the field as it attempts to align itself with
those more advanced (i.e. those using a Bayesian conclusion framework)
forensic disciplines.
FSCs are the most commonly performed task by forensic speech
scientists (Foulkes and French, 2012). The task of the expert is to provide expert
opinion on the speech evidence to the trier(s) of fact. The expert opinion in a
FSC is ideally presented in terms of the likelihood of obtaining the evidence
(corresponding to the similarities/differences between the criminal and suspect
samples) under the hypothesis that the samples came from the same person,
versus the probability of obtaining the evidence (the typicality of the analyzed
47
speech parameters) under the hypothesis that two different speakers produced
the criminal and suspect samples. The methodologies undertaken by an expert
in a FSC are varied, unlike DNA where the same techniques are routinely
applied across cases. That is to say that the methodologies involved in FSCs
need to be adapted on a case-by-case basis to varying extents. The reasons for
this are twofold: (1) speech data is complex in nature (confounding factors are
often present), and (2) there is no single speech parameter that is omnipresent
and can discriminate all speakers.
2.4.1 Complexity of Speech Data
Speech is inherently variable, so much so that phoneticians often make
reference to the simple fact that no two speech utterances produced even by the
same speaker are ever identical. It is this intra-speaker variation that sets
forensic speech science apart from some other forensic disciplines. DNA is an
example of forensic evidence where the criminal and suspect samples can be
identical. For speech, unlike DNA, it will never be the case that the probability of
obtaining the evidence given the prosecution hypothesis is ever equal to 1.
Variability within the speech of an individual can be caused by numerous
factors (e.g. the interlocutor, illness, speaking style, intoxication); however, the
maximum extent of variation that can be observed within a speaker is not
completely understood through currently available models in linguistics,
sociolinguistics, phonetics, or phonology.
The variation that is observed between speakers, or inter-speaker
variation, is also highly conditioned by both biological and anatomical factors
(e.g. vocal tract length, the rate at which vocal cords vibrate), as well as
48
phonological and social factors (e.g. age, sex, class; Chambers, 2005; Eckert,
2000; Wardhaugh, 2006). It is also possible for these variables to interact with
one another, whereby their effects are manifested in the speech of individuals
differently.
The levels of intra-speaker variation observed in speech recordings are
typically high; therefore, it is not surprising that many individual linguistic-
phonetic parameters analyzed in FSCs offer only small contributions to
advancing the task of speaker discrimination. For this reason, a forensic
phonetician will traditionally consider multiple phonetic-linguistic parameters
under a combined auditory and acoustic phonetic analysis (French and Stevens,
2013). As a result, the different phonetic-linguistic parameters in a FSC form
highly correlated systems and sub-systems owing to the relevant anatomical,
phonological, and social factors. The relationships that exist in the data when
multiple parameters are under consideration must be appropriately taken into
account in the evidence (as is also the case for other forensic sciences). This is
so that the conclusion presented in a FSC case is representative of the evidence,
and does not over- or under-estimate its strength.
Speech data present a number of challenges to phoneticians looking to
analyze phonetic-linguistic parameters in FSCs. This is because the probability
distributions associated with phonetic-linguistic parameters are variable. The
parameters can be discrete (categorical or qualitative, e.g. impressionistic
analysis of voice quality), continuous (e.g. formant frequencies), or a
combination of both (e.g. discrete at one level and continuous at another for the
same parameter). The continuous parameters are (as a convenient
simplification) traditionally assumed to be normally distributed. However, an
49
assumption of normality is not always advised as it can lead to miscarriages of
justice (e.g. believing a speaker to be an outlier when he/she is actually very
similar to the rest of the population). Algorithms used for calculating LRs (those
which assume normality in the data distributions) evaluate similarity and
typicality on the basis of data existing at all areas under the normal distribution
curve. If a normal distribution curve does not accurately describe the data, the
LR algorithms will compute similarity and typicality evaluations from
inaccurate descriptions of the data distributions. Additionally, it is possible that
the distribution of values of a parameter for an individual speaker is different
from the distribution of values of that parameter for a group of speakers.
Finally, in addition to the innate factors that make speech generally
complex is the inevitable reality that speech recordings made under forensic
conditions are often compromised in terms of quality. Criminal recordings are
increasingly recorded via cellular phones, and the recording/transmission
technologies involved may affect the quality of the recording. Telephone
bandwidth restrictions (Byrne and Foulkes, 2004; Enzinger, 2010b; Knzel,
2001), the distance (of the speaker) from the microphone (Vermeulen, 2009),
and cellular phone audio recording codecs (Gold, 2009) have been shown to
artificially attenuate portions of the speech signal, which in turn causes
unwanted changes to formant frequencies and the fundamental frequency.
Furthermore, criminal recordings are also susceptible to low signal-to-noise
ratios and high levels of background noise and/or overlapping speech. For this
reason, the task of extraction of the necessary parameters is made more difficult
in FSC analysis. Despite these problems being prevalent in criminal recordings,
they are typically not as severe (or not present at all) in direct and high-quality
50
recordings made in a police interviews (e.g. a recording of a suspect) or other
comparable situations.
Given that the suspect recording in a FSC case typically comes from a
police interview, there is often a mismatch in the conditions under which the
criminal and suspect samples are elicited. The criminal recording is frequently
made in situations that involve high emotional states, physical activity, or the
influence of drugs or alcohol. They also tend to be short in duration and limited
in content. This presents the forensic phonetician with additional complications.
2.4.2 The Phonetic Shibboleth
The search for the linguistic or phonetic shibboleth7 for discriminating
speakers has proved fruitless since research began in the field of FSS. This
should not come as a surprise, given the inherent complexity of speech data, as
outlined in 2.4.1. Research has shown that the vocal tract is highly plastic,
that no phonetic/linguistic parameter is omnipresent, that a phonetic/linguistic
parameter that makes one speaker different does not necessarily make another
speaker different, and that parameters that make a speaker differ can vary over
time. It is likely, furthermore, that it is the combination of parameters that
makes a speaker unique (Nolan, 1983; Rhodes, 2013; Rose, 2013a).
There is a large and growing body of literature devoted to identifying
phonetic and linguistic parameters that have ideal characteristics for FSC. These
criteria have been outlined by Nolan (1983, p. 11):
7Shibboleth is used here to refer to a single identifiable parameter that can discriminate
between all speakers.
51
1. High between-speaker variability: The parameter should show a high
degree of variation between speakers. If a single parameter cannot
show this then a set of parameters can be sought
2. Low within-speaker variability: The parameter should show
consistency throughout the speech of an individual, and be insensitive
to external factors (e.g. health, emotion, or interlocutor)
3. Resistance to attempted disguise or mimicry: The parameter must
withstand attempts on the part of the speaker to disguise his voice
4. Availability: Any parameter should provide an ample amount of data
in both the criminal and suspect samples
5. Robustness in transmission: The usefulness of a parameter will be
limited if its information is lost or reduced due to recording or
transmission technologies
6. Measurability: The extraction of the parameter must not be
prohibitively difficult
Criteria 3-6 are specifically concerned with practical issues that arise in
casework. Criterion 4 relates to the often limited amount of material an expert is
given to work with, while criteria 5 and 6 are associated with the recording and
transmission technologies typically used in forensic recordings. The first two
criteria suggested by Nolan (1983) identify the true difficulty of the FSC task.
That is, the expert has to identify and examine phonetic and linguistic
52
parameters that have high inter-speaker variation, but also low intra-speaker
variation. It is often the case that a phonetic or linguistic parameter meets a
single criterion, but a phonetic/linguistic shibboleth is yet to be found that
meets both criteria without exception.
2.4.2.1 Research Question 1
Given the difficulties and limitations in selecting highly discriminant
phonetic and linguistic parameters for analysis in FSCs, the most obvious
question is:
(1) What phonetic and linguistic parameters do practicing forensic
phoneticians (around the world) typically analyze in a FSC case and
which parameters do they view as being highly discriminant?
Before questioning the proper (or logically and legally correct) framework in
which to make conclusions about FSC evidence, it is important to consider the
methodologies, practices, and parameter selection that is involved in the actual
FSC analysis itself. Only after establishing the general expectations of the FSS
community should one begin to broach the problem of FSC conclusion
frameworks. For without any analyzed forensic speech evidence, there can be
no valid conclusion.
2.4.3 Current Conclusion Framework in the UK
The current practice for presenting FSC conclusions in a UK court is not
in the form of an LR as described in 2.2, but rather is that described in the UK
Position Statement that was introduced in 2007. The UK Position Statement was
motivated by concerns about the framework in which conclusions are typically
53
expressed in forensic speaker comparison cases (French and Harrison 2007 p.
137). The UK Position Statement stemmed from ruling of the Appeal Court of
England and Wales in R v. Doheny and Adams (1996), which showed that the
interpretation of the DNA evidence at the initial trial had been flawed by the
prosecutors fallacy (French and Harrison 2007). The introduction of the UK
Position Statement signified a shift in the role of the forensic phonetician when
presenting speech evidence. The foreword to the UK Position Statement
suggests that experts in the past were often trying to identify speakers (French
and Harrison, 2007, p. 138). However, under their new approach an expert
would not be making identifications per se. Instead, the expert will take on a
different role (not one of speaker identification), to provide an assessment of
whether the voice in the questioned recordings fits the description of the
suspect (French and Harrison 2007 p. 138). The UK Position Statement was
also proposed with the intention of aligning the field of FSC with more modern
thinking forensic sciences (French and Harrison 2007 p. 137).
The framework laid out in the UK Position Statement diverges from
previous FSC conclusions by offering a framework which involves a bipartite
assessment. The conclusion framework set out in French and Harrison (2007)
potentially involves a two-part decision. The first part concerns the assessment
of whether the samples are consistent with having been spoken by the same
person. The second part, which only comes into play if there is a positive
decision concerning consistency, involves an evaluation of how unusual or
distinctive the combination of features that are common to the samples may be.
An illustrated version of the UK Position Statement is provided in Figure 2.1.
54
Figure 2.1: Illustration of the UK Position Statement (Rose and Morrison, 2009, p.141)
The UK Position Statement is illustrated in Figure 2.1, where the decisions of
consistency and distinctiveness are serially ordered. A consistency decision has
three possible options: consistent, not-consistent, and no-decision. If a
conclusion about consistency cannot be made, then the expert concludes with a
single evaluation (i.e. not consistent, or no decision). In the event that the expert
finds the two speech samples to be consistent, s/he will then assess the
55
distinctiveness. The degree of distinctiveness is made on a five-point
impressionistic scale, ranging from not distinctive to exceptionally
distinctive. The assessment of distinctiveness in many cases must draw upon
the experience of the expert so that he/she can provide a statement of the
typicality of the criminal speech sample.
The UK Position Statement framework can be seen as a transitional
point, or a stepping stone, between a frequentist probability and an LR, where it
is not providing a single probability of the hypothesis (e.g. the speaker in the
criminal sample is likely to be person X), but not quite meeting the logical
framework of the LR. At first glance the judgments of consistency and
distinctiveness appear to mirror the numerator and denominator of an LR, as
the consistency and distinctiveness account for both the similarity and the
typicality of the speech recordings. However, the inner workings of the Position
Statement do not hold true to the logical framework of an LR. There are two
main reasons for this mismatch: (1) assessments are made on different scales,
and (2) there is no logical procedure for combining (and weighing) constituent
speech parameter evidence from a single case.
Rose and Morrison describe the assessment of consistency in the UK
Position Statement as being on a three-point scale (Rose and Morrison, 2009, p.
142). Although Rose and Morrison acknowledge that the decision about
consistency is categorical, one could argue that the assessment of consistency is
more accurately described as simply a ternary decision (rather than on a three-
point scale). This is due to the fact that the judgment is wholly categorical, and
the ternary decision cannot be intuitively placed on a scale. A scale would imply
some degree of hierarchy, and it is difficult to argue that, for example, no-
56
decision should be ranked before inconsistent (or vice versa). Therefore, the
assessment of consistency is discrete in nature and does not offer a gradient
assessment of the similarity (through quantification of the speech evidence), as
the numerical LR would ultimately provide. The assessment of distinctiveness is
on a scale of one to five; however, this does not follow the same logic as the
assessments of consistency. Thus, it is difficult to establish a working
relationship between the two assessments; instead they exist more as two
separate entities, where practitioners are trying to make a judgment on the
same piece of evidence.
The use of a five-point scale in the UK Position Statement makes the
framework prone to a cliff-edge effect (Aitken and Taroni, 2004). By imposing
defined boundaries an expert is faced with a hard decision. So, for example, if a
criminal sample has an F0 mean of 115 Hz, while the population mean is 90 Hz,
should the analysis of a speech sample lend itself to a distinctiveness
assessment of 3 (distinctive) or 4 (highly distinctive)? Should the boundary
between distinctive and highly distinctive be two standard deviations, or
perhaps three? Forcibly imposing categorical boundaries could potentially over-
or under-estimate the strength of evidence. Although the UK Position Statement
is susceptible to the cliff-edge effect, the same can actually be said for the verbal
LR scale provided by Evett (1998). Although the verbal scale suggested by Evett
(1998) is associated with Log10 LRs, the cliff-edge effect can still occur for those
Log10 LRs that lie close to the categorical boundaries.
The second inconsistency between the UK Position Statement and the LR
is the lack of a protocol for combining the strength of evidence of individual
phonetic-linguistic parameters. Under a Bayesian framework an expert is
57
expected to combine individual LRs for parameters that are mutually
independent (Kononenko, 1990) by multiplying their LRs. If an expert is to
navely combine correlated parameters without using appropriate statistical
weightings, s/he then runs the risk of over-or under-estimating the strength of
evidence (as s/he are essentially considering the same evidence multiple times).
Under the UK Position Statement, experts make assessments of consistency and
distinctiveness by informally considering all of the constituent pieces of analysis
together. As such, they are unlikely to adequately (or transparently) consider
the degree of correlation between the evidence. Therefore, conclusions made
under the UK Position Statement framework could over- or under-estimate the
strength of evidence.
Despite the disparity between the UK Position Statement and the LR, the
UK Position Statement possesses two highly attractive attributes. Firstly, it
allows the expert to avoid the undesirable and lengthy task of collecting
(quantitative, data-based) population statistics for all possible relevant
populations that could ever be required for a FSC case. Secondly, the framework
allows the expert to avoid the difficulty of calculating numerical LRs for all
phonetic-linguistic parameter distributions that do not fit into already existing
LR algorithms. These two attributes are technically part and parcel of the same
thing, as they together evaluate the denominator of the LR; however, the
modeling of phonetic-linguistic parameters is also pertinent to the numerator. It
is safe to argue that no LR algorithm could ever account for or encompass the
full complexity of speech data; therefore, perhaps the UK Position Statement is
right to circumvent fully quantitative population statistics and complicated
models for calculating LRs for all speech parameters. Through experience and
58
education, an expert is able to account for instances of accommodation, channel
mismatch, intoxication, emotional effects, and social factors. These factors tend
to manifest themselves differently in the speech of each individual speaker and
at different times. To create an algorithm that accounts for every individual, in
every instance, would be near impossible.
2.5 Likelihood Ratios in Forensic Speech Science
It is now generally accepted in forensics that the logically and legally correct
framework for expressing the results of forensic examinations is one in which
the output is a likelihood ratio (Saks & Koehler, 2005). The domain of forensic
speech science is no exception to this, and efforts have been made in the last
decade and a half (starting with Rose, 1999) to incorporate into the LR into
forensic phonetic- (and linguistic)-based research and casework. In forensic
phonetics, the LR essentially becomes a test of the similarity and typicality of
phonetic-linguistic parameters that are extracted from recordings. The
numerator of the LR contains the probability of obtaining the evidence given the
hypothesis that the speech came from the same speaker, while the denominator
is the probability of obtaining the evidence given the hypothesis that the speech
came from a different speaker (Rose, 2002). The same-speaker hypothesis is
determined by comparing speech parameters from the criminal and suspect
samples to establish the degree of similarity. The different speaker hypothesis is
determined by comparing speech parameters from the criminal speech sample
to those drawn from a relevant background population so as to establish the
degree of typicality. The probability obtained from the numerator is then
59
divided by the probability obtained from the denominator, and the result is the
LR for the given speech evidence.
The presence of LRs in FSC is largely confined to the research literature
with only a single case example of the LR being used in FSC casework to date
(Rose, 2012; 2013). The research carried out has predominantly had two main
foci: (1) assessing speaker discrimination using a numerical LR and (2) overall
improvements in LR methodologies. The next two sections ( 2.5.1 and 2.5.2)
provide an overview of methodological research that has been carried out, as
well as a review of the application of LRs in practice.
2.5.1 Likelihood Ratios in the Literature
This section focuses on the LR literature that investigates the use of
numerical LRs as a framework for carrying out the assessments of speaker
discrimination ability using phonetic and linguistic parameters, and the LR
literature that seeks to improve current methodologies.
2.5.1.1 Likelihood Ratios for Speaker Discrimination
The application of the LR framework to FSCs has focused almost
exclusively on vowels. Vowels can be easy to extract quantitative measurements
from, and so readily lend themselves to the calculation of numerical LRs. In
order to improve discrimination rates between speakers, researchers have
measured vowels with multiple methodologies: using mid-point formant values
(Alderman, 2004; Rose, 2007a; Rose, 2010a; Rose and Winter, 2010; Zhang et
al., 2008), formant trajectories of monophthongs and diphthongs (Atkinson,
2009; Enzinger, 2010a; Kinoshita and Osanai, 2006; Morrison, 2009a; Rose et
60
al., 2006), and long-term formant distributions over vowel mixtures (Becker et
al., 2008; French et al., 2012; Moos, 2010).
Comparatively speaking, non-vowel research for speaker discrimination
purposes has not been given the same amount of attention as vowel-based LR
research. Those studies that have been carried out on features other than
vowels have all focused on quantitative, multivariate data that is typically
normally distributed. The non-vowel parameters that have been investigated
include fundamental frequency (F0), voice onset time (VOT), nasals, laterals,
and fricatives (Kavanagh, 2010; 2011; 2013; Kinoshita, 2002; 2005; Kinoshita et
al., 2009, Coe, 2012). Traditional FSC does not just involve the analysis of vowels
and the non-vocalic features listed above. For this reason, further empirical
work is required to evaluate the discriminatory value of additional speech
parameters using a numerical LR.
2.5.1.1.1 Research Question 2
The current body of literature evaluating the discriminant ability of
speech parameters is plentiful. However, there are a number of speech
parameters that have not had their discriminant ability tested. To date,
parameters have been selected for analysis based principally on their ease of
measurement. Instead, it is proposed here that parameters should be selected
on the basis of their discriminatory merit as proposed on the basis of the
experience of forensic phoneticians. These considerations lead us to ask our
second research question:
61
(2) If experts are to provide their opinions on the most helpful speaker
discriminants are these selected parameters going to be good
speaker discriminants?
a. Furthermore, do experts expectations surrounding the
discriminant value of certain speech parameters match the
results of these parameters empirically-tested performance?
A simple hypothesis to test when empirically evaluating parameters that are
identified as being commonly used in FSCs and which experts propose to be
useful discriminants is that such parameters will perform better than speech
parameters selected for analysis arbitrarily (simply because they are easily
measurable and plentiful).
2.5.1.2 Improving Likelihood Ratio Methodologies
In addition to the LR literature that has assessed the discriminant ability
of phonetic parameters, there is a dedicated body of literature on
methodological advances in the calculation of LRs in other domains. In
particular, there have been methodological advances across a range of areas,
including the development of modeling techniques of data for calculating LRs
(Kinoshita, 2001; Morrison, 2011; Zhang et al., 2008), exploring the issues
surrounding correlated parameters (Gold and Hughes, 2012; Morrison et al.,
2010; Rose, 2006c; 2010b; Rose et al., 2004), identifying the relevant population
for the LR (Hughes, in progress; Hughes and Foulkes, 2012; Morrison et al.,
2012a; 2012b), exploring the amount of data that is preferred for the reference
population (Hughes, in progress; Hughes and Foulkes, 2012; Ishihara and
Kinoshita, 2008; Kinoshita and Ishihara, 2012), combining parameters
62
(Morrison et al., 2010; Morrison, 2013; Rose 2010a; 2010b, 2013a; Rose et al.,
2004; Zhang et al., 2008), system calibration (Morrison, 2012; Morrison et al.,
2010; Morrison and Kinoshita, 2008), and measures of system validity and
reliability (Morrison et al., 2010; Morrison and Kinoshita, 2008).
Despite these methodological developments, there are numerous
research questions relating to the calculation of LRs in FSCs that would greatly
benefit from further empirical study and assistance from forensic statisticians.
The majority of the previous research has perhaps neglected to acknowledge
the complexity of speech data and has opted for often convenient but erroneous
simplifications of basic linguistic principles in order to calculate LRs.
2.5.1.2.1 Research Question 3
As previously outlined in 2.4.2, Nolan (1983) recommended that a set
of parameters should be sought to show high between-speaker variability where
a single parameter alone is not sufficient. Given the large body of literature on
single speech parameters as discriminants and their limited discriminant
power, it is suggested that further work needs to heed Nolans suggestion.
(3) How well do speech parameters work in combination to discriminate
between speakers?
a. What steps need to be taken in order to appropriately
combine speech parameters?
b. Is the combination of multiple speech parameters always
better than individual parameters at discriminating between
speakers?
63
One would hypothesize that adding ever more parameters would further
advance the task of FSC, since theory and research tells us that speakers are
different from one another in a variety of ways. By combining multiple speech
parameters, it is proposed that a combined system will achieve better
discrimination performance than those achieved by single parameters.
2.5.2 Likelihood Ratios in Practice
The only publications that report the use of LRs for multiple speech
parameters in FSC casework are those of Rose (2012; 2013b) in connection with
a fraud case in Australia. The case of R v. Hufnagl (2008) revolved around a
large-scale telephone fraud of AUS$150 million, where a criminal sent a fax to JP
Morgan Chase bank, asking to transfer $150 million from the Australian
Commonwealth Superannuation Scheme to accounts in Switzerland, Greece, and
Hong Kong. Before the close of business, the criminal called the bank asking for
confirmation of the details in the fax he had sent. When the Australian
Commonwealth Superannuation Scheme realized their account was short by
$150 million, an investigation followed. A suspect was identified, and Rose was
asked to compare the recording of the fraudulent telephone call with recorded
telephone calls known to have been made by the suspect. The analysis and
report were produced five years prior to Roses publications about it (2012;
2013b), so he presents the original analysis that was carried out as well as a
retrospective critique of his analysis.
In the original analysis, he identified many tokens of the word yes in both
the criminal and suspect recordings, as well as the utterance not too bad in the
criminal recording and multiple occurrences of the same phrase in the suspect
64
recordings. Therefore, the majority of the analysis and the resulting LR were
based on phonetic/linguistic parameters measured from these words. In order
to establish the typicality of the criminals speech Rose defined the relevant
population to be adult male speaker[s] of General Australian English (Rose
2013b, p. 284). He then collected relevant speech samples from 35 adult males,
who served as the background population. The analysis of similarity was
comprised of formant measurements from /je/ in the word yes at three
designated time-points, the fundamental frequency (F0) in not too bad taken
from four designated time-points, categorical classification of high and low
tones in not too bad, formant measurements of the vowels in not too bad, and
the frequency cut-off in /s/ from the word yes (Rose, 2013b). After
(intentionally) navely combining the individual LRs from the parameters, an
overall LR (OLR) of around 11 million was calculated. Rose (2013b) explicitly
states that 11 million was an over-estimation of the strength of evidence, since
some degree of correlation had to exist between the parameters. For this
reason, parameters that were assumed to have some degree of correlation with
one another (e.g. formant measurements for certain vowels) were thrown out,
and a more conservative LR of 300,000 was reached.
Five years after the conclusion of the case, Rose provided a critique of the
analysis and presentation of the evidence under an LR framework. He notes a
number of developments made in the field since the R-v-Hufnagl case that could
have made a significant difference in his analysis. These include vowel (and
consonant) parameterization (e.g. formant dynamics; McDougall, 2004),
quantification of accuracy and precision (validity and reliability; e.g. Cllr and
EER for validity measures), and - most importantly - techniques to handle
65
between-parameter correlations for calculating OLRs (e.g. fusion). If any of
these developments were to have been implemented in R v. Hufnagl (2008), it
can confidently be said that the strength of the numerical LR would not be
identical to that presented in Rose (2013b; also shown through his reanalysis of
the case material); most likely, the strength of the LR would weaken as
correlated parameters were accounted for during the combination of speech
evidence (acknowledged in Rose, 2013b).
The final portion of Rose (2013b) commented upon the courts reaction
to the presentation of evidence in the form of a numerical LR, which is
something rarely discussed in forensic phonetics. The expert testimony did not
include a complete tutorial on the LR approach; rather, it offered a more
abstract presentation of the strength of evidence (the LR). Rose (2013b)
condensed his analysis into two main points for the jury, which he emphasized
on multiple occasions: (1) the LR is for estimating the strength of the evidence
and not the probability that the suspect is the criminal, and (2) the jury should
not give much weight to the specific value allocated to the LR, just that it was
very big. Whether Roses testimony made an impression on the triers of fact in
R-v-Hufnagl is unknown. However, the jury did return a guilty verdict (Rose,
2013b). Rose also notes that it was perhaps vital to his testimony that the judge
was encouraging towards his approach and that this helped him (Rose) to
articulate to the court the strength of the speech evidence. It can be assumed
that not all judges would act in the same manner, and presenting the same
testimony in front of a different judge might have been more challenging
without such support.
66
Overall, it is encouraging to see an example of a real case in which a
numerical LR framework was used. The introduction of Roses paper provides a
nice backdrop to the case and the type of speech material Rose chose to analyze.
The critique at the end of the paper is a positive contribution, as it shows how
the field has evolved in the past five years since the case analysis was
completed. The paper also shines light on the reception of the LR in a court,
which again often goes without attention in the literature. However, the paper
perhaps brings up more questions (both theoretical and practical) about the
implementation of the LR framework (as used by Rose) than it answers. For
instance, how does an expert begin to select parameters for analysis under an
LR framework? How can an expert argue why s/he has selected certain
parameters for analysis over other parameters? How is an expert to incorporate
qualitative/categorical parameters? And how many parameters need to be
analyzed to consider the evaluation to be complete?
Despite raising a new set of questions, Rose (2013b) makes three
pertinent statements with respect to LRs. These statements are particularly
relevant to the remainder of this thesis. The first is that real-world cases are
never the same and there is no one-size-fits-all with regards to methodology
(Rose, 2013b, p. 318). This means that the LR calculation is not the same in
every FSC case, or for every phonetic/linguistic parameter selected for analysis.
Therefore, the analysis that leads to an LR will always have to be adapted on a
case-to-case basis. The second statement asserted by Rose is that FSC might
lend itself more readily to a verbal LR over a numerical LR8. The reason for this
8A verbal LR is simply a verbal, rather than numerical, statement of the probability of obtaining
the evidence given the prosecution hypothesis over the probability of obtaining the evidence
67
is that precise figures may be misleading in that numerical LRs may be difficult
for the trier(s) of fact to interpret9 (Rose, 2013b, p. 305). The final statement
comes from Judge Hodgson (2002) but is reiterated by Rose (2013b): since not
all types of evidence in a trial can be sensibly assigned a LR there is no way of
mathematically combining la Bayes the LR-based evidence with the non-
numerically based evidence (Rose 2013b p. 316-317). This leaves one to
ponder whether there is really an explicit need for speech evidence to be
represented in numerical LR form. For example, would a phonetician ever be
able to quantify the exact tongue shape of a speakers //? In this instance, a
qualitative description of // will typically be more useful than a quantitative
one that is not completely transparent in its description. Should these types of
evidence always be unsuitable for expression in a numerical LR, will it be the
case that other phonetic-linguistic parameters can be made to fit the mold in the
form of LR algorithms that dictate specific quantitative forms? It is also
important to consider that if a numerical LR is used, only a partial assessment of
the speech evidence is feasible, given that numerical LRs cannot currently be
calculated for all speech parameters (because of the lack of appropriate
algorithms and/or the qualitative nature of certain parameters), and the lack of
population statistics in general.
2.5.2.1 Research Question 4
The literature review provided in the previous sections revealed a
number of limitations and difficulties that can occur when applying the
given the defense hypothesis. For example the verbal statement could be presented as it is
extremely more probable to obtain the given evidence under hypothesis x than y.
9 For example, is there really much of a difference between an LR of 1.1 x 10 14 and an LR of 1.11
x 1014?
68
numerical LR framework to FSCs, which are largely due to the complexity of
speech data. If the field is to continue in its efforts to align itself with more
advanced forensic disciplines (in terms of conclusion frameworks) that have
already adopted the LR framework (e.g. DNA), various aspects of the actual
calculation of an LR in a FSC should be reviewed and improved (e.g. modeling
techniques, population statistics, combining parameters for OLRs).
(4) For this reason, it is essential to ask: What are the practical
limitations/implications that need to be considered when using the
numerical LR framework in FSCs?
a. What recommendations, if any, can be provided following
attempts to implement the numerical LR framework?
b. What can a human-based (acoustic-phonetic) system tell the
field in respect of the ease with which a numerical LR can be
computed for FSCs?
The practical limitations and implications associated with the implementation
of a numerical LR will be discussed throughout this thesis. It is only through
empirical testing that these questions can be addressed.
2.6 Summary of Research Questions
This chapter has presented a series of research questions that have been
motivated by the prior literature and existing legal rulings with regard to
forensic evidence, while further developing the research aims of the thesis. This
section reiterates the research questions identified in this chapter, which will be
explored in the remainder of the thesis.
69
phoneticians (around the world) typically analyze in a FSC case
and which parameters do they recommend as being highly
discriminant?
(2) If experts are to provide their opinions on the most helpful
speaker discriminants will these selected parameters be good
speaker discriminants?
a. Furthermore, do experts expectations surrounding the

discriminant value of certain speech parameters match
these parameters empirically-tested performance?
(3) How well do speech parameters work in combination to
discriminate between speakers?
a. What steps need to be taken in order to appropriately

combine speech parameters?
b. Is the combination of multiple speech parameters always

better at discriminating between speakers? Are more
parameters better?
(4) What are the practical limitations/implications for using the
numerical LR framework in FSCs?
a. What recommendations, if any, can be provided following

attempts to implement a numerical LR framework?
b. What can a human-based (acoustic-phonetic) system tell

the field in regards of the ease with which a numerical LR
can be computed for FSCs?
70
Chapter 3 International Survey of
Forensic Speaker Comparison
Practices
3.1 Introduction
This chapter presents the results of the first comprehensive international
survey on forensic speaker comparison (FSC) practices.
The motivations for the survey were twofold:
(i) For the first time, to make available to the wider forensic, legal, and
speech science communities basic information concerning the working
practices of FSC experts around the world.
(ii) To draw upon the very considerable collective experience of FSC experts
worldwide in order to identify current working methods and features of speech
that are considered to have the greatest potential for discriminating between
individuals.
It will become apparent from the results presented below that there is a great
deal of variation in the methods of analysis, features selected for examination,
weighting attached to certain features relative to others, and frameworks used
for expressing the conclusions that arise from the comparisons. Some of the
differences found are, undoubtedly, a function of the rules, regulations and laws
of the institutions and jurisdictions in which the survey participants are
working. Others, however, would appear to be simply a matter of local tradition
or individual intellectual preference. The results are therefore discussed in the
71
context of the constraints on the admissibility of expert evidence in different
countries and are related to contemporary debates within forensic speech
science.
3.1.1 Background
While research has been carried out on many facets of FSS and FSCs,
there has not been any research that has comprehensively surveyed the FSC
practices employed by experts around the world. The extent to which the FSS
community had been aware of commonly used FSC practices has been limited to
the results of an exercise Cambier-Langeveld (2007) conducted using a fictional
FSC case. The objective of the exercise was not to survey practitioners, but
rather to observe and assess basic methods that participants chose to employ in
conducting the fictional FSC case. Cambier-Langevelds paper considers reports
from 10 of 12 participants based in 10 different countries. Her article reports on
some of the basic methods involved in a FSC case, which were confined to: the
length of recordings needed for speech samples, formant measurements,
fundamental frequency, and the formulations of conclusions. The results of the
exercise revealed inconsistencies in methods amongst the 10 participants.
However, it usefully relayed fundamental methodological information with
regard to FSCs that was previously unavailable.
The exercise conducted by Cambier-Langeveld (2007) was an attempt to
provide the field of FSS with a body of information relating to FSC methods, but
not to provide a wide-ranging picture of current practices. However, the study
did create interest and a platform on which to conduct further research into the
methodologies employed by the field for FSCs worldwide.
72
Hollien and Majewski (2009) discuss the prevalence of inconsistencies in
FSS practices with particular attention to FSCs, like those reported by Cambier-
Langeveld (2007). The authors suggest that the field of FSS lacks any real
consensus in terms of procedures and methods for FSC cases. They argue that it
is difficult to consider what level of scientific probability is robust enough to
determine the identity of a speaker, and that without any standards or common
practices it is difficult to make comparisons across different approaches. They
offer a protocol for a frequentist conclusion framework that they implement
(and that other experts could adopt should they wish to), which includes
confidence levels of their judgments. Despite their efforts to offer their own
standard and protocol for FSCs, the authors fail to acknowledge alternative
methodologies that are currently being implemented by experts in FSC cases
across the globe. I would argue that an understanding of the current state of the
field is a prerequisite for establishing any form of standards or protocols.
For a field that came to fruition in the late 1980s/ early 1990s (the time
at which acoustic and auditory phonetic analysis began regularly being used in
the UK courts at least, (French, p.c.)) little has been done to unify and
standardize the field over this time. While Cambier-Langeveld (2007) and
Hollien and Majewski (2009) argue that there is a lack of consensus in the field
of FSC and that standards are almost non-existent, I would suggest that the only
way to remedy such a fault is first to assess the current methodologies and
practices being used in FSCs by surveying expert forensic phoneticians around
the world.
73
3.2 The Survey
The survey was administered online using SurveyGizmo 3.0. It consisted of 78
questions related to all aspects of forensic speaker comparison casework. All
participants were kept anonymous and also given the option of answering some
or all questions. Although every question was answered by at least 30
participants, the variability in respondent numbers nevertheless dictates that
the majority of the results must be presented as percentages.
3.2.1 SurveyGizmo
SurveyGizmo is an online survey software tool for designing online
surveys, collecting data and performing analysis. [The] tool supports a variety
of online data collection methods including online surveys, online quizzes,
questionnaires, web forms, and landing pages (SurveyGizmo 2010). It was
selected as the medium for the survey over other similar websites for a number
of reasons: the server is secure, the package offers the ability to save and
continue (when taking the survey), an inexpensive Student Account with
enhanced privileges, and an excellent user interface for creating the survey. All
responses collected from the survey are saved on the SurveyGizmo server, and
answers can only be accessed by a username and password.
3.2.2 Methodology: Data Compilation
To complete the survey, participants were provided with a survey link in
an email invitation. They were then redirected to the SurveyGizmo website
where they gave their consent to participate in the survey and agreed to their
data being used in future research. After giving consent, participants were
provided with instructions (see Appendix A) as well as an outline of the survey
74
structure. They were allowed to stop the survey at any time and save it, so that
it could be completed at a later time. Many of the respondents took advantage of
this feature as the total time (including interruptions) it took most participants
to complete the survey ranged from 26 minutes to 64 hours.
Once all participants had submitted their answers to the survey
questions, the results were tabulated using Microsoft Excel.
3.3 Participants
Potential participants were contacted through their professional and research
organizations. Emails were sent to the European Network of Forensic Science
Institutes, the National Institute for Standards and Technology (NIST) for those
who participate in the NIST Speaker Recognition Evaluations, and the
International Association for Forensic Phonetics and Acoustics. Some
individuals working at government laboratories/agencies were contacted
through their employers. In total, 36 practicing forensic speech scientists
agreed to participate, and data were collected from July 2010 through March
2011.
3.3.1 Countries
Respondents (23 male; 13 female) were from the following 13 countries:
Australia, Austria, Brazil, China, Germany, Italy, the Netherlands, South Korea,
Spain, Sweden, Turkey, UK, and USA. Although the majority of the participants
were from Europe, a total of five continents were represented in the results.
75
3.3.2 Place of Work
Respondents identified their place of work 10 or affiliation. 18
participants represented universities or research institutes, followed by 13 who
were employed in government laboratories/agencies. 9 of the experts are
affiliated with private laboratories, and 7 work as individuals.
3.3.3 Experience
The total number of cases from respondents collective estimates was
18,221, ranging from 4 to 6,000, with a mean of 506. The respondents had a
range of 2 to 50 years of experience in FSC analysis, with a mean of 15.
3.4 Methods of Analysis
Participants responses showed that there is at present no consensus of opinion
in the scientific community as to how FSC analysis should be carried out.
Rather, a wide range of methods is employed. Methods may be grouped under
the following headings:
Auditory Phonetic Analysis Only (AuPA):
The expert listens analytically to the speech samples and attends to
aspects of speech at the segmental and suprasegmental levels.
Acoustic Phonetic Analysis Only (AcPA):
The expert analyses and quantifies physical parameters of the speech
signal using computer software. As with AuPA, this is labor-intensive,
involving a high degree of human input and judgment.
10 Some respondents are associated with multiple places of work.
76
Auditory Phonetic-cum-Acoustic Phonetic Analysis (AuPA+AcPA):
This combines the preceding two methods.
Analysis by Automatic Speaker Recognition System (ASR):
This requires the use of specialist software designed to estimate the
degree of similarity between speech samples based on statistical
models of features extracted automatically from the acoustic signal.
Such systems typically require minimal input from the analyst.
Analysis by Automatic Speaker Recognition System with Human

Analysis (HASR):
This involves the use of an automatic system in conjunction with
analysis of the auditory and/or acoustic phonetic kind. The survey
did not investigate the precise nature or extent of the auditory or
acoustic examinations experts used to supplement the ASR
component. The human-based supplementary analysis may range
from cursory holistic listening to detailed auditory and/or acoustic
examinations.
More detailed descriptions of these methods, either individually or relative to
one another, may be found, inter alia, in Baldwin and French (1990), Drygajlo
(2007), French (1994), French and Stevens (2013), Greenberg et al. (2010),
Jessen (2007a; 2008), and Knzel (1987).
The distribution of these methods across the 13 countries is provided in
Table 3.1.
77
Table 3.1: Methods of analysis employed by country
Method Countries
AuPA Netherlands, USA
AcPA Italy
Australia, Austria, Brazil, China, Germany,
AuPA+AcPA
Netherlands, Spain, Turkey, UK, USA
HASR Spain, Germany, South Korea, Sweden, USA
The distribution of the methods of analysis relative to type of workplace is
shown below in Table 3.2.
Table 3.2: Places of work against method of analysis employed
AuPA AcPA AuPA+AcPA HASR

university or research institute 2 1 13 3
government laboratory/agency 8 4
private laboratory 1 7 1
as an individual 7
As is evident in Table 3.2, the HASR method is used most frequently by
government laboratories/agencies (33% use it), as opposed to only 16% using
HASR in universities or research institutes. AuPA + AcPA is well distributed
across all places of work.
The specific features of speech that are analyzed and considered
important vary from analyst to analyst within each of the method categories.
The data relating to this variation are presented in 3.9.
3.5 Conclusion Frameworks
As with method of analysis, there is no consensus within the forensic speech
science community as to how conclusions are and should be expressed.
78
Currently, there is much debate in the field about the logical and legally
correct frameworks (French and Harrison, 2007; Rose and Morrison, 2009;
French et al., 2010).
A variety of frameworks for expressing conclusions is currently utilized
across the world. The conclusion frameworks may be grouped under the
following headings:
Binary Decision:
A two-way choice that either the criminal and suspect are the same
person or different people.
Classical Probability Scale (CPS):
The probability or likelihood of identity between the criminal and
suspect is stated. Typically, the assessment is a verbal rather than a
numerical one and it may use such terms as likely/ very likely to be
the same (or different) speakers. These types of judgments are often
labelled as frequentist.
Different probability scales are used by different experts,
causing concern amongst practitioners over the lack of clarity caused
by these different scales used for making conclusions (Broeders,
1999, p. 229). DNA evidence conclusions (presented using Likelihood
Ratios) in combination with the scientific status of forensic
evidence in the USA have had a large impact on all fields of forensics
in gradually undermining the traditional use of probability scales
(Broeders, 1999, p. 231). Conclusions made using CPSs do not
(intentionally or otherwise) incorporate any estimate of typicality,
79
and generally fail to acknowledge the defense hypothesis (e.g. the
evidence came from/was produced by someone other than the
suspect).
Likelihood Ratio (LR):
This expresses the results as the likelihood of finding the degree of
correspondence or non-correspondence between the samples on the
basis of the prosecution hypothesis (that they come from the same
speaker), against the defense hypothesis (that they come from
different speakers). Some analysts express the likelihood ratio as a
number; others do so verbally. Both verbal and numerical LRs
provide a strength of evidence statement (see 2.3.1 for more
information) in the form of a verbal or numerical conclusion,
respectively (Morrison, 2009).
Using LRs, unlike CPSs, allows for typicality assessments to
be made. This requires population statistics or a knowledge of the
population in question for a given piece of evidence. In light of the
paradigm shift (Morrison 2009) LRs are thought to be the most
logical way in which to express conclusions. Furthermore, the
National Research Council (NRC) report to Congress on
Strengthening Forensic Science in the United States recommends
Aitken and Taroni (2004), Evett (1990), and Evett et al. (2000), as
they provide the essential building blocks for the proper assessment
and communication of forensic findings (2009, p. 186). All three are
proponents of the likelihood ratio framework.
80
UK Position Statement:
This conclusion framework potentially involves a two-part decision.
The first part concerns the assessment of whether the samples are
compatible, or consistent, with having come from the same person.
The second part, which only comes into play if there is a positive
decision concerning consistency, involves an evaluation of how
unusual or distinctive the features common to the samples may be
(French and Harrison, 2007).
Expanded explanations of these various frameworks are to be found in,
inter alia, Broeders (2001), Champod and Evett (2000), French and Harrison
(2007), French et al. (2010), Jessen (2008), Morrison (2009c), and Rose and
Morrison (2009).
Some methods of analysis lend themselves more readily than others to
the adoption of certain conclusion frameworks. For example, some automatic
systems express the results of the comparison as a numerical LR as one of their
options. A breakdown of methods against conclusion frameworks appears in
Table 3.3.
Table 3.3: Methods used for analysis in forensic speaker comparisons against conclusion
frameworks
Binary Numerical Verbal UK Position

CPS Other
Decision LR LR Statement
AuPA 1 1
AcPA 1
AuPA + AcPA 2 10 1 2 10
HASR 3 2 1 1
81
As seen in Table 3.3, there is a tendency for participants using AuPA + AcPA to
adopt the classical probability scale and UK Position Statement conclusion
frameworks.
Table 3.4 breaks down conclusion frameworks by country. Some
countries appear more than once, as there were multiple respondents from the
same country, with individual experts implementing different conclusion
frameworks.
Table 3.4: Conclusion frameworks used by country
Conclusion Framework Countries

Binary Decision Brazil, China
Australia, Austria, Brazil, Germany,
Classical Probability Scale Netherlands, South Korea, Sweden,
UK, USA
Numerical LR Australia, Germany, Italy, Spain
Verbal LR Netherlands, USA
UK Position Statement Germany, Spain, Turkey, UK, USA
A Likert Scale was used to measure the respondents level of satisfaction with
the conclusion method s/he used. Likert ratings were averaged across
respondents. The scale ranged from 1 (extremely dissatisfied) to 6 (extremely
satisfied). Table 3.5 reports the number of experts responding, mean scores,
and standard deviations for satisfaction levels by conclusion frameworks.
Table 3.5: Satisfaction with conclusion framework
Mean Likert Standard Number of

Conclusion Framework
Rating Deviation Experts
Numerical LR 5.00 0.00 4
UK Position Statement 4.27 0.65 11
Verbal LR 4.00 0.00 3
Classical Probability Scale 3.69 0.95 13
Binary Decision 3.50 2.12 2
82
3.5.1 Population Statistics
Out of all respondents, 70% reported that they use some form of
population statistics in arriving at their conclusions. 58% stated that they had
personally collected population statistics for the incidence of occurrence of one
or more phonetic or acoustic features.
The features to which the statistics relate include fundamental frequency
(used by almost all of the 70%), articulation rate, voice onset time, long term
formant frequencies, and, where applicable, stammer/stutter. A number of the
respondents commented that if more population statistics were available they
would use them.
3.6 Guidelines
Respondents were asked whether they followed a protocol/set of guidelines in
each forensic speaker comparison case and if so whether they had been
involved in its design. 85% of respondents used some form of a protocol or set
of guidelines. For those following a protocol or guidelines in casework,
respondents were asked how their protocol/guidelines came into existence.
The responses are distributed in Table 3.6.
Table 3.6: Creation of guidelines/protocols
Origin Number
Developed personally 6
Developed in conjunction with colleagues 17
Given it by place of work 3
83
3.7 Casework Analysis: Alone or in Conjunction with Others
When asked whether they worked individually or in conjunction with
colleagues in carrying out speaker comparisons, participants provided the
information set out in Table 3.7.
Table 3.7: Workers involved in a single case
How Work is Carried Out Number

All work done individually 13
Work done with the help of an assistant 4
Work done in conjunction with other members of a group/team 10
Work is done individually and checked by someone else 7
3.8 Casework in Foreign Languages
Besides carrying out casework in their native language, 56% of experts stated
that they also conduct casework in other languages. Collectively, these experts
have worked on cases in over 40 different languages other than their own.
Of those who work with other languages, 94% require the assistance of a
native speaker of the language in question, and, of those requiring such
assistance, 56% deem it necessary for the assistant to have a qualification in
linguistics and/or phonetics.
3.9 Features Examined in Detail
This section reports on the aspects of recorded speech that respondents take
into account or consider important in FSC cases.
3.9.1 Phonetic Features
Respondents were asked whether and with what frequency they
examined the following features.
84
3.9.1.1 Segmental Features
All respondents analyze vowel and consonant sounds in the course of
their examinations. With regard to vowels, 81% invariably carried out some
form of analysis and 13% routinely did so. 94% of all experts evaluated the
auditory quality of vowels, 97% carried out some form of formant examinations
and 58% measured vowel durations.
Of those undertaking formant examinations, all measure the second
resonance (F2). 87% of respondents reported measuring F1 and an equal
percentage reported measuring F3. 17% of respondents stated that they
measure F4. Only 10% of respondents measuring formants measured F1-F4,
63% measured F1-F3, and 10% measured either F1 and F2 or F2 and F3. In
respect of which aspects of formants are examined, 94% reported measuring
center (i.e. temporal midpoint) frequencies of formants of monophthongs, 71%
reported measuring formant trajectories of diphthongs and 45% examined
vowel-consonant or consonant-vowel formant transitions. 35% stated that they
examine formant bandwidth and 13% reported examining formant densities.
In relation to consonants, all respondents reported subjecting them to
some form of examination; 52% invariably did so. For all experts, 88% of
respondents reported evaluating auditory quality. 82% stated that they
examined aspects of timing and 48% reported measuring the frequencies of
energy loci.
Table 6 reports the frequency with which consonants, broken down by
manner of articulation, are analyzed in FSC cases. Respondents gave their
answers using a 6-point Likert Scale ranging from 1 (never) to 6 (always). The
number of experts responding, mean Likert ratings, and standard deviations are
85
represented in Table 3.8 for those respondents who are native English speakers
(and working on English cases only).
Table 3.8: Frequency of consonant analysis in English
Manner of Mean Likert Standard Number of

Articulation Rating Deviation Experts
Fricatives 4.85 1.21 13
Plosives 4.73 1.49 11
Approximants 4.50 1.27 10
Laterals 4.46 1.13 13
Nasals 4.08 1.24 12
Affricates 3.82 1.47 11
Taps/Flaps 3.70 1.77 10
Trills 3.18 2.04 11
3.9.1.2 Suprasegmental Features
All respondents (excluding those using AuPA only) routinely measure
fundamental frequency in their comparisons. With respect to what they
measure, for those conducting some form of AcPA, 94% reported measuring the
mean, 41% the median, 34% the mode, 72% standard deviation, 25% the
alternative baseline (the value (in Hz) that falls 7.64% below the F0; Lindh,
2007), and 6% measure the range. Considering all aspects of fundamental
frequency listed above, the most common combination for analysis was the
mean plus the standard deviation (22% of participants), followed by 19%
examining mean, median, mode and standard deviation together. Only 9%
measure the mean, median, mode, standard deviation, and alternative baseline.
Single respondents also reported measuring the coefficient of variation, also
known as the varco (the standard deviation divided by the mean (Jessen et al.,
2005)), the first and third quartiles, and kurtosis/skew. It is important to note
that although many respondents reported analyzing the fundamental
86
frequency, a large proportion point out that it is usually of little help. One
respondent stated that fundamental frequency is] usually used as an
elimination tool rather than an identification tool.
94% of the respondents who include an AuPA stated that they examine
voice quality as part of their overall procedure, although only 77% of these
invariably or routinely examine it. Further to this, 61% of those who examine
voice quality do so using a recognized scheme (e.g. Laver, 1980) or modified
version of such a scheme, for its description. Of those experts examining voice
quality the large majority (63%) reported using the Laver Voice Profile Analysis
Scheme (VPAS) or a modified version of it. 21% of experts perform an auditory
analysis of voice quality and provide some form of a verbal description. The
remaining experts use the GRBAS scheme (Grade, Roughness, Breathiness,
Asthenia, Strain; see Bhuta et al. 2004 for more information) or a modified
version of it (13%), and a single expert (3%) reported using LTS spectra (long-
term spectra) for examining voice quality. Furthermore, three experts provided
insightful commentary regarding the discriminant power of voice quality,
voice quality can often be central to the analysis and is best analyzed
systematically using a detailed scheme such as the Edinburgh VPA. Another
respondent states that voice quality is frequently strongly discriminating in
forensic speaker comparisons and the third expert comments that they are
increasingly of the view that voice quality is one of the most valuable but least
well understood parameters.
85% of all respondents stated that they examine intonation with one or
another level of frequency. However, of these only 25% look at intonation
87
invariably. The specific aspects of intonation vary, with tonality 11 being
reported more than tonicity12, 67% vs. 38% of respondents (Ladd, 1996, p. 10).
Tails of tone units were examined by 46% and heads by 29%.
93% of respondents stated that they analyze tempo with varying
degrees of frequency. Of those analyzing tempo, 81% apply a formal measure
(e.g. speaking rate (SR) or articulation rate; Knzel, 1997). For formal measures,
articulation rate (AR) was reported most frequently by 47% of respondents
compared to only 19% that use speaking rate and 16% that use both
articulation rate and speaking rate. Those using AR were asked how they
defined a syllable, and 93% of the respondents reported using phonetic
syllables for AR rather than linguistic ones. 73% stated that they examine
speech rhythm with varying regularity.
3.9.2 Non-Phonetic Features
3.9.2.1 Higher-Order Linguistic Features
In addition to examining phonetic features, 76% of all respondents
reported examining discourse features and/or conversational behaviors
(discourse markers, aspects of turn-taking, telephone opening and closing
behaviors, patterns of code switching). 88% of all experts stated that they
examine lexico-grammatical usage. Lexical features were examined most
frequently, followed by syntax and morphology.
11 Tonality marks one kind of unit of language activity and roughly where each such unit
begins and ends one tone group is as it were one move in a speech act (Halliday 1967 p. 30).
12 Tonicity marks the focal point of each such unit of activity: every move has one (major), or
one major and one minor, concentration point, shown by the location of the tonic syllable, the
start of the tonic (Halliday 1967 p. 30).
88
3.9.2.2 Non-Linguistic Features
94% of the respondents who answered this question reported examining
non-linguistic features at least some of the time. In descending frequency order,
specific features were as follows: filled pauses, tongue clicking13, audible
breathing, throat clearing, and laughter.
3.10 What is Considered Discriminant
In addition to being asked about features within the linguistic, phonetic and
acoustic domains, respectively, participants were given the opportunity to
identify which feature from any domain they found most useful for
discriminating speakers. For all respondents together, voice quality was
reported most often (32%), followed by dialect/accent variants and vowel
formants (both 28%). 20% reported speaking tempo and fundamental
frequency as useful parameters. This was followed by rhythm (16%). Lexical
and grammatical choices, vowel and consonant realizations, phonological
processes (e.g. connected speech processes) and fluency were all reported by
13% of the respondents. One respondent went as far as stating that vowel
formant analysis is rarely insightful.
Interestingly, though perhaps not surprisingly, the vast majority of
participants alluded to the fact that despite some individual parameters having
significant weight, it is the overall combination of features that they consider
crucial in discriminating between speakers. In Aristotelian terms, The whole is
greater than the sum of the parts (Aristotle, Metaphysica 10f-1045a).
13It is perhaps better to classify tongue clicking as a linguistic feature when it is used in a
inherently functional way, such as that described in Chapter 7. However, at the time of the
survey, clicks were classified as non-linguistic.
89
3.11 Discussion
The purpose of this chapter has not been to advance any argument or to
develop theoretical propositions. Rather, its objective has been the much more
mundane one answering to the motivations set out in 3.1: laying out basic
factual information concerning the practice of FSC internationally in the present
day and drawing upon the collective expertise of FSC experts worldwide so as
to identify current working methods and features of speech that are considered
to have the greatest potential for discriminating between individuals.
Those not directly involved in this specialist field but working, for
example, in other aspects of phonetics or linguistics, may well be surprised at
the lack of consensus over such fundamental matters as how speech samples
are to be analyzed and compared, which aspects of the samples are to be
assigned greatest importance during the analytic process, and how conclusions
are to be expressed at the end of it. However, we are assured by those working
in various other fields of forensic science that the level of dissensus uncovered
by the present survey is by no means unique to forensic speaker comparison.
Indeed, some of the practices and preferences found here are undoubtedly
dictated or constrained by the rules of the institutions and firms in which the
participants work. Where those organizations include other forensic science
disciplines, the options, particularly for the framing of conclusions, may be laid
down unilaterally for all types of casework investigation undertaken under
their auspices. For instance, the Dutch government forensic science facility, the
Netherlands Forensic Institute, requires that the outcomes of every
investigation undertaken by its employees, irrespective of the forensic
discipline, be expressed within a Bayesian likelihood ratio framework (Meuwly,
90
p. c.). Likewise, one of the participants in the present study who used a binary
decision format when expressing conclusions stated that to do so was a
requirement of his/her employer.
Over and above the rules laid down by public and private sector
laboratories, nations may also set down requirements, either by statute or via
case law. Some jurisdictions, notably the England and Wales division of the UK,
have been extremely non-prescriptive in this respect, according the expert a
very high degree of autonomy and discretion over the methods of analysis
he/she adopts and the way the outcomes are formulated. In respect of forensic
speaker comparison evidence, the England and Wales position was affirmed in
the Appeal Court ruling R -v- Robb (1991), in which the court ruled that
whether or not an expert used any acoustic testing was entirely his/her own
decision, and re-iterated in relation to the same issue in the more recent appeal
R -v- Flynn and St John (2008). Indeed, the main analytic issues over which the
higher courts have seen fit to pass down general prohibitions to forensic
experts concern the use of statistics in representing the strength of evidence (cf.
R -v- Doheny and Adams (1996); R -v- T (2011)). Where experts enjoy freedom
of choice, one might expect their preferences to be influenced by individual
intellectual commitments. However, in spite of the latitude allowed by the UK
legal system, it is of note that all nine UK experts taking part in the study use the
combined AuPA + AcPA method. Indeed, this method is the predominant one
across all countries represented in the survey (25 = AuPA + AcPA; 10 = other
see Table 3.3). Thus, although the results show a wide range of variation in
methods, there is nevertheless a very large degree of convergence.
91
As for the future, certain trends can be predicted. One is that as time
goes on and further improvements are made to the error rates of
(semi)automatic systems and to their capabilities for handling real case (i.e.,
non-studio) recordings, one would expect to find such systems increasingly
being incorporated into casework alongside the AuPA + AcPA approach. This
development would be particularly apposite in the USA, where the appeal court
ruling Daubert -v- Merrell Dow Pharmaceuticals Inc. (1993) is taken by many
lower courts as the benchmark for admissibility of expert evidence, and within
that ruling is the statement that the court ordinarily should consider the known
or potential error rate of the method. ASRs readily lend themselves to meeting
this criterion, and, indeed, many systems are subject to such testing as part of
the annual NIST evaluations (Greenberg et al., 2010). Further, a number of
ASRs have an LR as one of their easily selectable options for representing the
results of speaker comparisons. As seen in Table 3.3, most experts currently
express their conclusions in terms of a classical probability scale (14), whilst
only half as many (7) use some form of LR (3 = verbal LR; 4 = numerical LR).
The increasing use of ASR software, together with the current paradigm shift
in forensic science towards Bayesian reasoning, and the use of LRs for
presenting results, would lead one to expect an increase in the number of
experts using LRs and a corresponding decrease in the use of other conclusion
frameworks.
Morrison (2009c, p. 298) suggests that today we are in the midst of
what Saks and Koehler (2005) have called a paradigm shift in the evaluation
and presentation of evidence in the forensic sciences which deal with the
comparison of the quantifiable properties of objects of known and questioned
92
origin, e.g., DNA profiles, finger marks, hairs, fibres, glass fragments, tool marks,
handwriting and voice recordings. However he fails to acknowledge the fact
that not all speech evidence is of the quantifiable type, as demonstrated in the
survey results from the present study. The Bayesian framework of likelihood
ratios has been adopted by many fields in the forensic disciplines where
quantifiable evidence is of the norm and qualitative evidence is something that
does not necessarily come into question (e.g. DNA or fingerprints). It is
important to recognize that speech does not consist entirely of measurements.
There are elements of speech that are best described/analyzed qualitatively (i.e.
certain aspects of voice quality (e.g. lingual body orientation), lexical, syntactic,
or morphological choices, audible breathing, laughter). If such features can be
quantified in some form, then it is plausible that we will one day see an entire
forensic speaker comparison case completed in a Bayesian framework, but until
then there will still be experts who will continue to present such features in a
qualitative form, whether that is alongside a LR conclusion or another form of
conclusion (e.g. CPS or UK Position Statement).
Additionally, in light of the popularity of the Bayesian framework, it can
be predicted that more research on LRs will be carried out. This can be seen as a
positive trend, as parameters that experts found to be discriminant in their
experience (as reported in this survey), may now be tested empirically, and
general strength of evidence statements can potentially be attributed to certain
features. Given this, experts and researchers in the field of forensic speech
science can give appropriate weight in forensic casework to those features
93
found to be most discriminant through intrinsic14 and extrinsic15 likelihood
ratio testing.
Finally, those differences that currently exist across practitioners may be
reduced through blueprints16 and drives for international co-operation and
cross-border transferability of forensic science evidence (e.g. House of
Commons Northern Ireland Affairs Committee, 2009). And, of course, the
prerequisite for resolution of differences is knowledge of their existence.
Insofar as the present study lays bare that information, it may be considered to
be making a modest first step towards international unity.
3.12 Limitations
The International Survey on Forensic Speaker Comparison Practices has three
limitations. The first is the limited number of experts who took part in the
survey. Ideally, one would like to work with a larger sample size in order to
represent the total population of forensic speech scientists as accurately as
possible. Thirty-six is a large proportion of practicing forensic phoneticians.
However, it would have been preferable to include even more forensic
phoneticians and to have been able to represent a greater number of countries,
languages, and methods in order to achieve the most accurate representation of
current practices in forensic speaker comparison.
The second limitation is the lack of representation from those experts
using ASR alone. As is evident in the NIST speaker recognition evaluations
14 Intrinsic likelihood ratio testing uses the same set of speakers (e.g. from the same speech
corpus) for both the test and reference samples.
15 Extrinsic likelihood ratio testing uses different sets of speakers (e.g. from different speech
corpora) for the test and reference samples.

16 Blueprint is used here to refer to some form of standards documents.
94
(Campbell et al., 2009), there are a number of experts around the world who
use ASR alone, and the survey results presented in this chapter fail to represent
this fact. Despite efforts to recruit participants that utilize ASR alone, no such
experts responded to the survey.
The final limitation is the simple fact that these results have a limited
shelf life meaning that the field is always changing and forever evolving, and
these results are only a snapshot of the field as it stood in 2010-2011. The
trends seen in the survey will certainly vary in the future as more research is
carried out and new methodologies are put into practice.
3.13 Parameters Chosen for Further Analysis
As stated in 1.3, this survey served in part as a hunting ground for identifying
the speech parameters believed by experts to hold the greatest discriminant
potential. Based on responses from the practitioners, I now identify four
parameters from the survey that experts found to be highly discriminant and/or
analyzed relatively often in casework: articulation rate, long-term formant
distributions, long-term fundamental frequency, and clicks (velaric ingressive
stops). The subsequent chapters analyze each of the four parameters in turn,
while referring to the discriminant expectations of a given parameter.
95
Chapter 4 Articulation Rate
4.1 Introduction
Forensic phoneticians have suggested that speech tempo is an important
parameter for forensic speaker comparisons, with 93% of experts analyzing
speech tempo and 73% of those doing so with varying regularity (Chapter 3). It
is also reported that when asked which parameters they found highly
discriminant in forensic speaker comparisons, 20% of all experts reported that
they found speech tempo to be the most useful parameter for discriminating
speakers. Overall, speech tempo was ranked as the third most helpful
parameter (alongside F0) of all possible parameters used in a forensic speaker
comparison. Analyzing speech tempo in detail for a large, homogeneous group
of individuals provides insight into the distribution of and variation within the
parameter and thereby its ability to discriminate between speakers.
In forensic phonetics, speech tempo is typically quantified as either
speaking rate (SR) or articulation rate (AR; Knzel, 1997). Both speaking and
articulation rate measure the speech tempo of an individual, but the two
measures capture slightly different aspects of tempo. Speaking rate (SR) can be
defined as the rate of speech of the whole speaking-turn. It therefore includes
all speech material (linguistic or non-linguistic), together with any silent
pauses, that are contained within the overall speaking-turn (Laver 1994 p.
158). Articulation rate (AR) is defined as the rate at which a given utterance is
produced. The speech material measured by articulation rate therefore
excludes silent pauses by virtue of the definition of an utterance, which begins
and ends with silence (Laver 1994 p. 158). The difference between speaking
96
rate and articulation rate is that the former includes disfluencies and
filled/unfilled pauses in the calculation, whereas the latter excludes disfluencies
and unfilled pauses. Within the field of forensic speech science the majority of
experts report a preference for measuring articulation rate rather than
speaking rate in forensic speaker comparison casework (Chapter 3).
Population statistics for articulation rate on a large, homogeneous scale
(100+ speakers) exist for German and Chinese, but as yet, there has not been a
similar study carried out on English. This study presents the analysis for the
ARs and standard deviations (SD) of 100 Southern Standard British English
(SSBE) male speakers. The results concern both the inter-speaker and intra-
speaker variation of AR, as well as assessing the evidential value of AR as a
parameter in forensic speaker comparisons.
4.2 Literature Review
Articulation rate has been investigated in British English in small-scale studies.
Goldman-Eisler (1956) was one of the first to analyze and calculate articulation
rate in a population. In her study, she examined the spontaneous speech of eight
British adults in 30- to 60-minute interview-type recordings. AR was calculated
by counting the number of syllables (the definition of the syllable and interval
type were undefined in the study) in an utterance, with an utterance defined as
periods of speech lasting from a preceding question or utterance of an
interviewer to the next, which is usually occasioned by the subject having come
to a natural stop or pause (Goldman-Eisler, 1956, p. 137). It was found that the
mean AR across speakers was 4.95 syllables per second, ranging from 4.4 for
97
the slowest to 5.9 for the fastest speakers. The mean standard deviation across
speakers was 0.91 syllables per second (syll/sec), ranging from 0.54 to 1.48.
Kirchhbel and Howard (2011) also collected articulation rate figures
for British English, while investigating properties of speech that could
potentially be correlated with emotional/psychological stress. The study
examines the spontaneous speech of a loosely homogeneous group of 10 young
Southern British males in mock police interviews. Along with the AR and SR
results for the psychologically-stressed speech of the subjects, Kirchhbel and
Howard (2011) also provided baseline results for the speakers using interpause
stretches to obtain AR measurements. The mean AR for speakers was 5.81
syll/sec with a range of 5.14 to 7.00 and a standard deviation of 0.89 syll/sec
(range 0.79 to 1.01).
In a later study, Goldman-Eisler (1968) further examined AR as well as
SR. However, this study focused on intra-speaker variation. It was observed that
AR exhibits fairly little intra-speaker variation, whereas there is much more
variability present in SR. Henze (1953) conducted a similar investigation in
German using spontaneous speech in the form of story-telling elicited by a film,
and the same observations were made. It is noted that different speech tasks,
for example read versus spontaneous speech produced in different emotional
states, can cause differences in the pauses that speakers use (e.g. number, kind,
duration), in turn causing variations in speech rate across different speaking
tasks. Articulation rate differed slightly across the different tasks, but it was
found to be relatively stable across tasks in these two studies.
With respect to the implications for forensic speaker comparison, Knzel
(1997) examined AR, SR, and various pausing parameters in German. He
98
retested claims that intra-speaker variation was lower in AR than SR. He
confirmed prior results showing that intra-speaker variation is much smaller.
For the experiment, the read and spontaneous speech of five males and five
females was analyzed, and SR was found to be higher in read speech than in
spontaneous speech. This is largely due to the fact that speakers use far fewer
hesitation pauses (i.e. filled and unfilled) in read speech than in spontaneous.
AR, on the other hand, was not significantly different between read and
spontaneous speech, and AR for individual speakers had coefficients of variance
that were smaller than they were with SR. To further evaluate the possible
discriminating power of SR and AR, Knzel looked at cumulative distributions
of both intra- and inter-speaker differences. According to the equal error rates
(see 4.6.1) calculated, AR was found to have more speaker-discriminating
power than SR.
Following Knzels (1997) conclusion that AR is a better discriminator
than SR, investigators have begun to examine AR in more detail. In keeping with
Knzels conclusion that AR will have to be interpreted with caution when used
in forensic speaker recognition until its possibilities and limitations have been
assessed on the basis of genuine case material and large numbers of speakers
(1997, p. 79), additional studies have been conducted in both German and
Chinese. Jessen (2007b) analyzed the AR of 100 male speakers of German. AR
was measured for both spontaneous and read speech. It was found that, contra
Knzel (1997), the mean AR was significantly higher in read speech. Overall,
Jessen found the mean AR for the 100 speakers was 5.21 syll/sec. In order to
calculate ARs, he was the first to implement a new methodology in which
memory stretches (Jessen, 2007b, p. 53) were utilized rather than interpause
99
stretches and intonation phrases (Trouvain 2004, p. 50), which had been
commonly used in previous studies. Jessen describes the methodology behind
memory stretches as the phonetic expert going through the speech signal
and [selecting] portions of fluent speech containing a number of syllables that
can easily be retained in short-term memory. After listening several times the
expert then counts the number of syllables that he/she is able to recall from
memory to be included in this portion of speech (Jessen, 2007b, pp. 54-55).
Cao and Wang (2011) followed the methodology of Jessen (2007b) and
examined the ARs for 101 male Mandarin Chinese speakers in spontaneous
telephone speech. They investigated inter-speaker and intra-speaker variation
in AR, and found both the global ARs (GAR) and means of local ARs (LARmean)
to be fairly normally distributed (GAR and LAR mean are explained further in
4.3.2). The mean global articulation rate (GAR) was 6.58 syll/sec and the mean
of the local articulation rates (LARmean) was 6.66 syll/sec. They also report
that the range of AR for a given speaker is relatively small and stable. However,
ARs in Mandarin Chinese appeared to be higher than English and German
studies. The authors attribute the difference to the simpler syllable structure
found in Chinese. Chinese syllables are largely /CV/ in shape; therefore more
syllables per second can be produced than is possible with the inherently longer
syllables in German and English (Cao and Wang, 2011, p. 398).
Although AR has been examined in large-scale studies of both German
and Chinese, the greatest number of subjects examined in a previous study of
English speakers is 50, and many studies examined are based on considerably
fewer. Table 4.1 provides an overview of AR studies conducted on English.
100
Table 4.1: Overview of articulation rate studies
Study Subjects Task AR mean avg. AR range SD mean avg. SD range

Goldman-Eisler (1956) British: 8 subjects Interviews: Spontaneous 4.95 4.4-5.9 0.91 0.54-1.48
American: 20 male and 20

Robb et al. (2004) Rainbow Passage: Read 5.27 0.40
female adults
New Zealand: 20 male and 20
Rainbow Passage: Read 5.70 0.47
female adults
Doherty & Lee (2009) Read (1st time through
Irish: 22 males 5.68
Rainbow Passage)
Read (2nd time through
Irish: 22 males 6.05
Rainbow Passage)
Irish: 22 males Conversation: Spontaneous 5.88
Read (1st time through
Irish: 22 females 5.38
Rainbow Passage)
Read (2nd time through
Irish: 22 females 5.67
Rainbow Passage)
Irish: 22 females Conversation: Spontaneous 5.58
Jacewicz et al. (2009) North Carolina: 50 adults Conversation: Spontaneous 5.41 0.48
North Carolina: 50 adults Sentences: Read 3.27 0.44
Wisconsin: 44 adults Conversation: Spontaneous 4.81 0.54
Wisconsin: 44 adults Sentences: Read 3.54 0.34
Kirchhbel & Howard (2011) SSBE: 10 males Interviews: Spontaneous 5.81 5.14-7.00 0.89 0.79-1.01
101
As can be seen in Table 4.1, a number of studies have examined AR for both
read and spontaneous speech, but to date only two small-scale studies on
British English have been carried out (Goldman-Eisler, 1956; Kirchhbel and
Howard, 2011). Combined, the results for ARs from these in respect of
spontaneous speech have a mean rate of 5.29 syllables per second, with the
slowest mean at 4.81 syll/sec and the fastest at 5.88 syll/sec. It is important to
note that these figures are the result of studies of only a few varieties of English,
and how other varieties and dialects may pattern is unknown. The most recent
AR study on British English (Kirchhbel and Howard, 2011) has a difference of
more than 1.00 syll/sec relative to Goldman-Eislers (1956) study carried out
about 55 years earlier. It is hypothesized that the results of the present study
will pattern more closely with those of Kirchhbel and Howard (2011) than
those of Goldman-Eisler (1956), as the former is based on a more
demographically and linguistically homogeneous group of speakers and uses a
similar methodology to the present study (the methodology in Goldman-Eisler
(1956) is not transparent and is therefore difficult to compare).
4.3 Population Statistics for Articulation Rate
The following section presents the collection of population statistics for
articulation rate in a large, homogeneous group of 100 male speakers. This data
serves as the first of its kind in providing detailed information on the
distribution of and variation in articulation rate for a large group of individuals
who speak SSBE.
102
4.3.1 Data
The data for the current chapter as well as subsequent chapters come
from the Dynamic Variability in Speech database (DyViS) recorded at the
University of Cambridge (Nolan et al., 2009). The DyViS database is a large
speech corpus collected under simulated forensic conditions (de Jong et al.
2007). It is comprised of recordings of 100 male speakers of Southern Standard
British English (hereafter referred to as SSBE) aged 18-25. This group of
speakers is meant to represent a homogeneous population in respect of sex, age,
and accent group. All speakers were recorded under both studio and telephone
recording conditions for Task 2 (see below), and under studio recording
conditions for a number of different speaking styles (i.e. Task 1, Task 3, and
Task 4). The DyViS recordings include four tasks identified below (adapted
from de Jong et al., 2007):
Task 1: simulated police interview (studio quality)

Task 2 telephone conversation with accomplice (studio and telephone quality)
Task 3: reading passage (studio quality)
Task 4: reading sentences (studio quality)
The first task in DyViS is a simulated police interview, whereby the
speaker is interrogated in a mock police investigation in relation to a (fictional)
drug trafficking crime. The speech is spontaneous insofar as speakers were
given visual stimuli (e.g. pictures of people and places) to prompt the
construction of their responses to the investigator (interlocutor). There were a
number of target words from the visual stimuli that were elicited by the
interlocutor (i.e. the interlocutor asked the speaker specific questions in order
to elicit the target words). The second task in DyViS is a telephone conversation
103
between the speaker and his accomplice Robert Freeman (the interlocutor is
the same person for all 100 speakers). Task 2 is recorded from the studio end of
the telephone call as well as via an intercepted external BT landline. The second
task, like Task 1, is spontaneous speech, whereby the interlocutor questions the
speaker in a mock police interview. The interlocutor for Task 2 elicits from the
speaker the same target words as those used in the police interview, which
inevitably leads the discussion in Task 1 and Task 2 to be very similar.
Tasks 3 and 4 of DyViS are both forms of read speech. Task 3 consists of
a read news report pertaining to the alleged drug trafficking crime. The same
target words are included in the read report. Task 4 is read speech from
controlled sentences that have a large number of SSBE vowels in nuclear non-
final position (i.e. in closed syllables), with six repetitions each.
The studies carried out in the remainder of this thesis will include only
data from either Task 1 or Task 2 (studio quality, spontaneous speech) of the
DyViS database. The current chapter uses Task 2 for calculating articulation
rate.
The DyViS studio recordings were all made using a Marantz PMD670
portable solid state recorder at a sampling rate of 44.1 kHz and 16 bit depth (de
Jong et al., 2007). All speakers were recorded via a Sennheiser ME64-K6
cardioid condenser microphone positioned approximately 20 cm from the
speakers mouth. The recordings were made in a sound-treated room in the
Phonetics Laboratory at the University of Cambridge (Nolan et al. 2009:40).
104
4.3.2 Methodology
The Task 2 recordings used for the current study were 15 to 25 minutes
in duration. However, only the relevant amount of material was analyzed in
order to extract between 26 and 32 memory stretches. 26 was chosen as the
lower boundary for memory stretches because it was the maximum number
of tokens that could be extracted from the shortest of the 100 recordings. The
upper boundary of 32 was chosen semi-arbitrarily (only to have a large number
of tokens for calculating likelihood ratios).
The general methodology employed in this study follows very closely
that of Jessen (2007b). In measuring AR a number of decisions have to be made
(Knzel, 1997; Trouvain, 2004). As Jessen (2007b, p. 53) explains, the first
concern is the kind of linguistic unit on the basis of which AR is counted. As
noted in Chapter 3, the majority of forensic phoneticians use the syllable as a
unit of measurement, rather than sound segments or words, in turn producing
AR rates in syll/sec as opposed to words per second (or minute). As a native
speaker of a language, one has a fairly reliable intuitive ability to count the
number of syllables in a specific stretch of speech. In terms of analysis, this
avoids the need to become involved in examining intensity peaks in the acoustic
signal on a syllable-by-syllable basis. Jessen also mentions that the syllable is
probably more a cognitive/linguistic unit grounded in the physics of speech
production (Keating 1988, cited in Jessen, 2007b, p. 53). For these reasons,
syllables in this study were determined auditorily through careful listening.
The second important decision for the measurement of AR relates to
whether one should define syllables phonologically or phonetically. A
phonological syllable is defined in terms of the lexicon and grammatical rules
105
of the language whereas a phonetic syllable is one that is manifested in
phonetic reality (Jessen 2007b p. 53). Jessen gives an example using the
phrase did you eat yet? Phonologically we would count this as having four
syllables; however, in reality the number of syllables may be reduced or in some
rare cases even increased. If the phrase were to be reduced it might be realized
as perhaps two syllables as in jeet yet (Jessen 2007b). For this reason the use
of phonetic versus phonological syllables makes a difference in terms of AR
counts. In a case where a phrase is phonetically only two syllables, AR will
obviously be lower than if the same phrase was counted as four phonological
syllables (see Jessen, 2007b, pp. 53-54 for further discussion). Jessen (2007b)
suggests that syllables are best defined phonetically, rather than phonologically.
This is because often in casework, speaker comparisons include speech from
different dialects or foreign accents, and in certain cases it might be difficult to
determine what the phonological form should be (Jessen, 2007b). However,
Jessen (2007b) also notes that:
[C]ounting actual syllables can lead to curious artefacts when a
speaker in speaking rapidly deletes [phonetic] syllables, whereas
another speaker might reduce or delete perhaps the same number
of [phonetic] sounds but still preserves the number of underlying
syllables. In such a case the former speaker ends up with lower AR
than the latter although both would be about equally fast if AR
were based on canonical rather than actual syllables.
Given that the present study is based on recordings of a linguistically
homogeneous population with the same accent and that counting syllables
106
phonetically has been shown to cause curious artefacts17 (Jessen, 2007b, p.
54; Koreman, 2006), the current study is based on phonological syllables rather
than phonetic syllables.
The third methodological decision, and the one which is perhaps the
most influential on the results, involves the kind of speech interval that is
selected for determining AR. The AR can be calculated for the entire duration of
fluent portions in a recording. This number is known as global AR.
Alternatively, by taking multiple fluent speech stretches within a recording,
local ARs can be calculated (Jessen, 2007b, p. 54). Miller et al. (1984) showed
that speakers often change their speech tempo over the course of longer
utterances. Therefore, in order to capture such changes in tempo that may
occur within a single recording, it is more useful to obtain local ARs. Previously,
researchers have commonly used interpause stretches and intonation phrases
to identify speech intervals over which to calculate local ARs (Trouvain, 2004).
Following Jessen (2007b), in order to avoid possible empirical or
methodological problems associated with the two aforementioned methods of
selecting speech intervals, a much simpler and more pragmatic approach was
chosen for this study. Interpause stretches tend to result in intervals that are
extremely variable in length due to pausing behaviors (which might reintroduce
the influence of pausing which AR tries to eliminate; Jessen, 2007b). Intonation
pauses are reliant on phonetic and linguistic judgments made by the analyst,
which result in variation of the interval lengths depending on the experts
interpretation (Jessen, 2007b). Therefore, the speech interval used in the
17These artefacts occur when one speaker may be speaking quickly and as a consequence
deletes phonetic syllables, whereas another speaker is typically inclined to reduce or
completely delete the same number of phonetic sounds. However, this speaker is able to
preserve the number of underlying phonological syllables.
107
current study for computing local ARs is referred to as a memory stretch
(Jessen, 2007b, p. 54). After listening several times to that interval of speech,
the expert then counts the number of syllables that he/she is able to recall from
memory being included in this portion of speech. Three examples of memory
stretches from Speaker 036 are presented in the Table 4.2.
Table 4.2: Examples of memory stretches for Speaker 036
Number of Time (in

Memory Stretch
Syllables seconds)
(a) I defended you gallantly 8 0.984
(b) They wanted to know about a car park 10 1.367
(c) They didnt elaborate or anything 11 1.161
Sony Sound Forge Audio Studio 10.0 was used for analysis. Speech intervals
were only selected at least two minutes into the recording, to allow the speaker
to become comfortable speaking to his accomplice in the presence of the
recording equipment. Like Jessen, speech intervals containing fluent speech
were chosen, and the region marked out. After listening several times, I would
type out the speech phrase on the region marker tag. Following this, I would
count the number of syllables included in that interval. After collecting enough
memory stretches, it was possible to view all recorded regions that listed the
number of syllables and included the length of the speech segment. Those
figures were entered into Microsoft Excel and local and global ARs as well as
standard deviations were computed for all speakers.
The procedure described above was applied to all 100 recordings
analyzed in this study. The mean and standard deviation of AR for each speaker
are used for analysis and are reported in 4.3.2 below. The maximum number
of syllables in a memory stretch was 26, but most stretches contained between
108
7 and 11 syllables (in order not to push the limits18 (Jessen, 2007b, p. 55), and
avoid mistakes following Jessen (2007b)). In line with the methodology
implemented by Jessen, four syllables or more per stretch were used. The
threshold is in place in order to avoid the inclusion of very short interpause
stretches that could unduly increase the effect of phrase-final lengthening on
the calculated articulation rate (Jessen 2007b, p. 55). It is important to note
that each memory stretch consisted of only fluent speech (for speech intervals).
Fluent speech was defined as speech that did not include the following: any kind
of pauses, either filled or unfilled, repeated syllables, unintelligible speech, and
any syllable lengthening (judged subjectively) that went beyond canonical non-
hesitation durations in English. The mean number of memory stretches
measured per speaker was approximately 30, with a standard deviation of 2.1, a
range of 26-32, and 2,993 total ARs calculated for the 100 speakers.
4.3.3 Results
The distributions of the local AR means and the standard deviations for
individuals are presented in Figures 4.1 and 4.2. The y-axis represents the
number of speakers that fall within a given range and the x-axis depicts
articulation rate presented as syllables per second.
That is in order to avoid trying to remember such an extensive interval of speech that
18
mistakes are made when trying to recall it, as this could potentially affect the resulting ARs.
109
Mean Articulation Rates
18
Number of Speakers 16
14
12
10
8
6
4
2
0
Articulation Rates (syllables/second)
Figure 4.1: Distribution of mean articulation rates
There is a roughly normal distribution19 for the mean ARs, as illustrated in
Figure 4.1. The mean AR for the population is 6.02 syll/sec, with an overall
range of mean AR from 4.57 to 7.79 syll/sec. The standard deviation of the
mean is 0.64 syll/sec. The 100 speakers have mean ARs within a 3.22 syll/sec
window.
The data were checked for two levels of outliers. This thesis defines
suspected outliers as falling between 1.5 times the interquartile range (IQR)
and 3 times the IQR, plus or minus the first or third quartiles. Any outliers that
fall outside the upper bounds of 3 times the IQR are confirmed as definite or
extreme outliers in this thesis. The mean AR has six suspected outliers at 7.23
19 Normality was judged visually, and not through statistical testing.
110
syll/sec, 7.24 syll/sec, 7.47 syll/sec (x2), 7.53 syll/sec, and 7.79 syll/sec.
However, there were no extreme outliers for mean AR.
Standard Deviation for Articulation Rates

25
Number of Speakers
20
15
10
Standard Deviation (syllables/second)
Figure 4.2: Distribution of standard deviations in articulation rate
The standard deviations for AR within speakers appear normally
distributed. The mean SD is 1.20 syll/sec, with a range of 0.72 and 3.95 syll/sec.
Those speakers who lie towards the left end of the x-axis are considered
relatively more consistent in their AR than those speakers who fall towards the
right end, who are characterized as having a more variable AR. The SDs of the
100 speakers lie within a range of 3.23 syll/sec, which is a larger range (by 0.01
syll/sec) than the range of means found for AR (see Figure 4.1). AR has three
suspected outliers at 1.72 syll/sec, 1.77 syll/sec, and 1.87 syll/sec. There are
also two extreme outliers at 2.36 syll/sec and 3.95 syll/sec.
111
The cumulative distribution graph of means in Figure 4.3 below shows
the percentile within which a given AR falls. The y-axis is the cumulative
percent of the population, and the x-axis represents AR.
Mean Articulation Rates

100
90
Cumulative Proportion (%)
80
70
60
50
40
30
20
10
0
4.4 4.6 4.8 5.0 5.2 5.4 5.6 5.8 6.0 6.2 6.4 6.6 6.8 7.0 7.2 7.4 7.6 7.8
Articulation Rate (syllables/second)
Figure 4.3: Cumulative percentages for mean articulation rate
The curve in Figure 4.3 is characterized by a steep central portion, but rather
gentle gradients at both ends. 1 SD from the mean gives a range of
approximately 5.3 and 6.6 syll/sec, into which roughly 73% of the population
falls. The cumulative distribution of individuals SDs is illustrated in Figure 4.4,
which follows the same template as that of Figure 4.3.
112
Standard Deviations for Articulation Rate
100
Cumulative Proportion (%) 90

80
70
60
50
40
30
20
10
0
0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.4 4.0
Standard Deviation (syllables/second)
Figure 4.4: Cumulative percentages for standard deviation in articulation rate
The curve in Figure 4.4 is similar to the curve seen in Figure 4.3, as it is
characterized by a steep central portion and gentle gradients at both ends.
However, the slopes at the ends are not as gradual as the curve of the mean
distributions (Figure 4.3). 1 SD from the mean SD gives a range from
approximately 0.82 to 1.58 syll/sec, a band in which roughly 84% of individuals
SDs fall.
Comparing the intra-speaker variation to the inter-speaker variation, it
is important to note that the mean SD (1.2 syll/sec) for a speaker is about twice
the SD (0.64 syll/sec) for between-speaker variation. One would be more likely
to find higher levels of variation within any given speaker of SSBE than between
that speaker and others. This variability is also shown in the variance ratio,
which is a calculated by dividing the squared between-speaker SD by the mean
squared within-speaker SD (Rose et al., 2006). A value of less than one indicates
113
that there is more variation within speakers than between them. A value
greater than one indicates that there is more variation between speakers than
within speakers. The variance ratio for AR is 0.2844, which confirms that there
is more variation within individuals than there is between them.
4.3.4 Discussion
In addition to providing population statistics, there are three main points
to be drawn from the results reported in 4.4.0. The first is that the ARs found
in the current study are very different from those found by older AR studies,
namely Goldman-Eisler (1956), Robb et al. (2004), and Jacewicz et al. (2009).
Goldman-Eisler (1956) reported a mean of 4.95 syllables per second, a mean
range of 4.40 to 5.90, and a standard deviation of 0.91 syllables per second
(range = 0.54 to 1.48). Her research was based on the spontaneous speech of
eight British adults recorded in 30- to 60-minute interviews. Her method of
calculating AR permitted certain disfluencies to be included in the material (e.g.
unnatural sound prolongations), and this perhaps in part accounts for her lower
mean AR than that found in the present study. Another reason for higher AR
results in the present study could also be due to the use of phonological
syllables rather than phonetic syllables. The use of phonetic syllables could
potentially lead to lower ARs, as it counts only those syllables which are actually
articulated by the speaker (see 4.3.2 for the example of did you eat yet (4
syllables) versus jeet yet (2 syllables)).
The second point is that claims that forensic practitioners who took part
in the survey reported in Chapter 3 made about AR being a useful speaker
discriminant appear to be misguided, as AR is a weak discriminator, because
114
there is more variation occurring within speakers than between them. For this
reason, discriminating between individuals is difficult when a person has a
typical mean AR. However, this is not to say that the parameter is not helpful
when discriminating between speakers with lower or higher mean ARs.
The final point is that forensic phoneticians need to take care when using
AR or SR in forensic speaker comparison analysis, since SR is even more
variable within a speaker than between speakers (Knzel, 2007). This
parameter is best used in combination with other parameters for discriminating
between speakers, but may carry more weight when AR is used to discriminate
individuals who fall near the outer boundaries of the distribution.
4.3.5 Limitations
A possible limitation of the present study is the selection of memory
stretches as the speech intervals over which syllables are to be counted.
Choosing a speech interval is dependent on the short-term memory of the
analyst calculating the AR. This can potentially lead to high levels of variation in
AR when measured by different analysts. Ideally, forensic methodologies should
be robust and easily replicable across many analysts in order to achieve
comparable results. For this reason it is important to verify the AR results
calculated from memory stretches by comparing memory stretch interval
results to those obtained from more commonly defined and objective interval
(i.e. inter-pause stretches).
4.4 Redefining the Speech Interval
The following section compares the results found for memory stretches in 4.3
to those found for inter-pause stretches.
115
4.4.1 Methodology
Twenty-five of the same speakers as those reported on above were
randomly selected from Task 2 of DyViS, and five minutes of speech from each
individual was analyzed starting at a point two minutes into each speech
sample. It is important to note that for comparison purposes, the mean AR for
an individual using memory stretches was only calculated using intervals from
the same five-minute speech sample as was used when calculating AR using
inter-pause stretches. The remaining aspects of the methodology were also kept
consistent, and syllables were defined phonologically.
Inter-pause stretches are defined here as both filled and unfilled pauses
that lasted 130ms or longer (Dankoviov 1997) but were not stop closures.
The interval also had to include at least five syllables. The criteria for items that
were excluded from analysis in an interval were identical to those set by the
exclusion rules in 4.3.1. This meant that intervals excluded any kind of pauses,
either filled or unfilled, repeated syllables, unintelligible speech, and any
syllable lengthening that went beyond the phonological requirements of
English. A mean of 40 intervals was measured per speaker, with a range of 26-
58, and amounting to 1,011 ARs in total. The mean number of syllables per
interval was 9.97, with a range of 5 to 37 syllables across all speakers.
4.4.2 Results
The mean ARs for both memory stretches and inter-pause stretches are
presented below in Figure 4.5. The y-axis represents the AR in syll/sec and the
x-axis shows the 25 randomly selected speakers from the 100 speakers in the
DyViS Database.
116
Mean ARs for 25 Speakers: Memory Stretches vs. Inter-pause
6.8
Stretches
Memory Stretches
Articulation Rate (syllables/second)
6.6 Inter-pause Stretches
6.4
6.2
6.0
5.8
5.6
5.4
5.2
5.0
16 19 24 25 30 31 35 39 50 55 56 57 60 62 63 66 67 68 71 75 76 78 85 87 91
Speaker
Figure 4.5: Comparison of mean articulation rate for memory stretches versus inter-pause
stretches
Figure 4.5 provides mean ARs generated using both methodologies. The means
for each speaker are displayed above one another to allow for a visual
comparison of the differences found between them. The two AR methodologies
prove unpredictable in terms of indicating a trend for whether one
methodology produces consistently higher or lower ARs than its counterpart, as
evident in the crossing lines in Figure 4.5. All speakers show relatively small
differences (especially speakers 031, 050, 056, 063, 075, 078, and 085) in their
mean ARs. However, some speakers have larger differences (e.g. speakers 016,
035, 068, 071, 076, and 091) than others. Using the absolute values of the
differences, the average (mean) AR difference across the 25 speakers is 0.286
syll/sec, with a range of 0.001 to 0.75 syll/sec.
117
The mean AR for the 25 speakers using inter-pause stretches was 5.98
syll/sec, compared to the mean AR of the same speakers calculated with
memory stretches, which was 5.96 syll/sec. The mean AR calculated by the two
different methodologies differs by two-hundredths of a second. This is a minute
amount, given that the mean ARs of the speakers are between 5.00 and 7.00
syll/sec and the mean SD for the 25 speakers (using memory stretches) was
0.64 (syll/sec). Using a Wilcoxon signed rank test for the two sets of data, the
null hypothesis is retained (there is no significant difference between the two
methods), as the p-value is 0.74. This provides a validation of the memory
stretch method for the calculation of ARs.
4.4.3 Discussion
The present study gives rise to two important conclusions. The first is
that AR results appear to be unaffected by the definition of the speech interval
as long as the following are kept consistent: the basic unit of speech defined
here as the phonological syllable, and the exclusion rules. Based on the findings
in 4.4.2 and the experience gained from calculating 125 mean ARs, I am now
of the opinion that mean AR measurements are affected more by the exclusion
rules than they are by the actual definition of the speech interval (memory
stretch vs. inter-pause unit). The exclusion rules were described in 4.3.2 and
concern what speech can be excluded from analysis, e.g. whether false starts,
unnatural prolongation, and unintelligible speech are to be included or
excluded. These exclusion rules can vary from analyst to analyst, as one might
find a repetition such as I-I-I-I am going to the store should be excluded, but
something such as I am I am going to the store should be included. The
118
judgments made with respect to exclusion in an analysis may be exercised by
the analyst at a level below his/her conscious awareness. However, speaking
from experience, an analyst remains relatively consistent working within the
constraints of the rules. Therefore, the differences that were found among the
mean ARs for the 25 speakers could more likely be attributed to variability that
naturally occurs when taking AR measurements (eg. the precise start/stop
times of intervals), and acoustic measurements in general (Harrison, 2004),
rather than to a difference caused by the definition of a speech interval (i.e.
memory stretches vs. inter-pause stretches).
The final conclusion to be drawn here is that reaffirmation that the
results found in 4.3 using memory stretches are valid and reliable
measurements, since there was no significant difference found between the
results arrived at using methodologies that incorporate memory stretches
versus inter-pause stretches. Furthermore, the amount of time it takes to
calculate ARs using inter-pause stretches is far greater than the time it takes to
calculate ARs using memory stretches. Therefore, analysts might be advised to
use memory stretches rather than inter-pause stretches in order to use time
more efficiently but still calculate reliable results.
4.5 Manipulating the Syllable Requirements
Given that the methodology for selecting speech intervals does not appear to
affect AR figures significantly, it is helpful to consider the effects that the
syllable requirements of a speech interval have on AR figures. This section
reports on the manipulation of the number of syllables used in a speech
119
interval. AR was recalculated for 25 speakers using different syllable lengths for
speech intervals.
4.5.1 Methodology
The data for the first 25 speakers of the study that were presented in
4.3 were used to recalculate mean ARs using different minimum syllable
requirements for speech intervals. There were seven possible minimum syllable
requirements for the speech interval, ranging from four to ten20. Microsoft Excel
was used to remove tokens from the speakers data for each given requirement
and mean ARs and SDs for all individuals were recalculated once tokens had
been removed from each minimum syllable level.
4.5.2 Results
The mean AR and SD results across all speakers for different minimum
syllable requirements in a speech interval are provided in Table 4.3. The table
details for each minimum syllable level the mean number of tokens included for
calculating speakers overall AR as well as the groups mean AR, and the groups
mean SD.
Table 4.3: Summary of articulation rate statistics when varying the minimum number of
syllables in the speech interval
Mean Number of Mean AR Mean SD of

Syllables
Tokens (syll/sec) Speakers
4< 29.96 5.738 1.134
5< 29.04 5.762 1.120
6< 26.96 5.844 1.105
7< 24.40 5.929 1.089
8< 20.92 6.024 1.072
9< 17.12 6.096 1.072
10 < 13.32 6.129 0.973
20This range was chosen based on the number of tokens available for the different syllable
lengths.
120
There is a prominent trend present in Table 4.3: as the minimum number of
syllables required for a speech interval increases, the mean AR for speakers also
increases while the mean SD decreases. This shows that perhaps AR becomes
slightly more stable within individuals as the minimum syllable count per
speech interval is increased. However, it is necessary to examine the results for
changes in within-speaker variation as well as changes in between-speaker
variation. Table 4.4 shows the mean differences in mean, SD, and the difference
() for speakers at different minimum syllable requirement levels. The first
column indicates the minimum number of syllables required for a speech
interval and the second and third columns display the mean for mean AR and
SD. The s in columns two and three are calculated by taking the average value
(mean AR or SD) at a given syllable level and subtracting the baseline means
(those values calculated for 4< syllables). Positive values represent an increase,
while a negative value would indicate a decrease in AR.
Table 4.4: Within-speaker differences for articulation rate
Within-speaker
Mean for Mean Mean for SD
AR (syll/sec) (syll/sec)
5< 0.024 -0.013
6< 0.106 -0.029
7< 0.192 -0.045
8< 0.286 -0.062
9< 0.359 -0.062
10 < 0.392 -0.160
Table 4.4 shows that individuals are patterning similarly to the group as a
whole in that mean AR increases and SD decreases as the minimum number of
syllables in a speech interval is increased. It appears that the higher the
121
minimum number of syllables in a speech interval the more stable a speakers
AR becomes. This could also be due in part to the decreasing number of tokens
involved in the calculation at the 10< syllable level. However, the number of
tokens is still rather robust for syllables at the nine or more level and below,
across which there is still a demonstrable decrease in within-speaker variation.
Table 4.5 examines between-speaker differences found for different minimum
syllable requirements. The first column indicates the minimum number of
syllables required for a speech interval and the second presents the mean for
the SD of the mean ARs. As in Table 4.4, a positive value shows an increase,
while a negative value shows a decrease in AR.
Table 4.5: Between-speaker differences for articulation rate
Between-speaker
Mean for SD of AR Means
(syllables/second)
5< -0.022
6< -0.004
7< -0.002
8< 0.014
9< -0.006
10 < 0.028
The overall trend displayed in Figure 4.8 is inconsistent, in that as the minimum
syllable requirement is increased the SD of the AR means fluctuates (both
increases and decreases to different degrees). Across all minimum syllable
requirement levels the mean is 0.001 syll/sec. This means that the variation
between speakers mean ARs remains relatively stable despite the changes
made to the minimum number of syllables required in a speech interval.
122
4.5.3 Discussion
The conclusion to be drawn from the results obtained by increasing the
minimum number of syllables required in a speech interval is that within-
speaker variation will decrease while the between-speaker variation remains
rather unchanged. This is important for the use of AR in an approach utilizing
likelihood ratios and could potentially lead to stronger evidential values for AR
as a discriminant. In the next section, LRs are calculated for AR, as well as to
ascertain whether increasing the minimum number of syllables in a speech
interval improves the strength of evidence for AR.
4.6 Likelihood Ratios
LRs provide a framework for the estimation of the strength of evidence under
competing defense and prosecution hypotheses (see Chapter 2). In order to
assess the discriminatory power of AR numerically, same speaker (SS) and
different (DS) LRs are calculated and plotted.
4.6.1 Methodology
The LR calculations for AR were performed using a MatLab
implementation of Aitken and Lucys (2004) Multivariate Kernel-Density
(MVKD) formula (Morrison 2007). MVKD was chosen over Lindleys (1977)
univariate LR, because it can account for inter- and intra-speaker variation (i.e.
using kernel-densities) whereas Lindleys (1977) formula makes LR
calculations based on normally-distributed data (i.e. a single distribution curve)
for both the LR numerator and denominator. Morrison also suggests that
because Lindleys LR formula cannot account for occasion-dependent within-
123
speaker variation (Morrison 2008 p. 97) the MVKD approach is more suitable
for analyzing speech evidence, although it is not ideal.
The MVKD formula provided by Aitken and Lucy (2004) assumes that
within-speaker variability is normally distributed (numerator). The between-
speaker variation, however, is not assumed to be distributed normally and is
estimated using kernel-density, a measure which accounts for skewed
distributions. The Gaussian mixture model - universal background model
(GMM-UBM21; Reynolds et al., 2000) has also been proposed for calculating LRs
in ASR, and has been used to calculate LRs in phonetic/linguistic-based FSCs
(Becker et al., 2008; French et al., 2012; Rose and Winter, 2010). GMM-UBM,
like MVKD, can accommodate multivariate data; however, GMM-UBM models
the data differently from MVKD. GMM-UBM utilizes GMMs to characterize
distributions instead of kernel densities (as per MVKD). The most significant
difference between GMM-UBM and MVKD is that the MAP (maximum a
posteriori; Reynolds et al., 2000) background model for GMM-UBM is person-
independent and is compared against a model of person-specific parameter
characteristics when comparing same (SS) and different (DS) speaker pairs
(Reynolds et al., 2000). Morrison has found that a GMM-UBM, which does not
assume normal distribution for within- or between-speaker values, performed
both better and worse than MVKD on different occasions (Lindh and Morrison,
2011; Morrison, 2011). However, the application of the GMM-UBM has
predominantly been tested using automatic systems. Despite these findings,
21GMM refers to the way in which the data are modeled. GMM is a parametric density function
that is comprised of a number of component functions (Gaussian; Reynolds et al., 2000). UBM
refers to the way in which the background population in modeled. A UBM is used to represent
general, person-independent parameters to be compared against a model of person-specific
parameters (Reynolds et al., 2000).
124
MVKD has been shown to provide reliable and important strength-of-evidence
results across many studies using more traditional acoustic-phonetic features
(Hughes, 2011; Morrison, 2008; 2009a; Rose, 2006a; 2007a, 2007b), and will
therefore be applied here to AR.
A MatLab script (ss_ds_lrs.m)22 was used to run multiple same-speaker
(SS) and different-speaker (DS) LR calculations for AR. The script calls for the
100 speakers samples to be split in half (i.e. 50/50), such that SS comparisons
may be performed (50 SS comparisons), which in turn results in 2,450 DS
comparisons (50*49). Speakers 001-050 act as the speaker comparisons, while
speakers 051-100 act as the background population. The calculated raw LRs
were transformed using natural and base10 logarithms. The transformation,
allows zero, rather than one, to act as the center point between the support for
Hp and Hd. The log transforms are also beneficial in normalizing distributions
which may be skewed by large and infrequent values.
The magnitudes of LRs are discussed and assessed with reference to
Table 4.6. The verbal scale, adapted from Champod and Evett (2000), is based
on identifying a Log10 LR that corresponds to a verbal expression representing
the strength of evidence in favor of the prosecution hypothesis (Hp: same
speaker) or the defense hypothesis (Hd: different speakers).
22 The script was developed by Phil Harrison from J P French Associates.
125
Table 4.6: Expressions for strength of evidence in terms of Log10 LR and the corresponding
verbal expression following Champod and Evetts (2000) verbal scale
Log10 LR Verbal expression

<-4 Very strong evidence to support Hd
-4 to -3 Strong evidence to support Hd
-3 to -2 Moderately strong evidence to support Hd
-2 to -1 Moderate evidence to support Hd
-1 to 0 Limited evidence to support Hd
0 to 1 Limited evidence to support Hp
1 to 2 Moderate evidence to support Hp
2 to 3 Moderately strong evidence to support Hp
3 to 4 Strong evidence to support Hp
>4 Very strong evidence to support Hp
This scale was previously used by the UK Forensic Science Service (Champod
and Evett, 2000), and will serve as a means of strength-of-evidence evaluation
for the remainder of the thesis.
Performance of the system is discussed in respect of log-LR cost (Cllr)
and equal error rate (EER), which are both metrics of system validity. The term
validity refers to how well a system (in this thesis a system can be an
individual parameter or multiple parameters combined) can distinguish
between same-speaker (SS) and different-speaker (DS) pairs. Severity of
performance error was assessed using Cllr, which is a common assessment used
in automatic speaker recognition/comparison (Ramos Castro, 2007). The Cllr is
a Bayesian error metric that quantifies the ability of the system to output LRs
that align correctly with the prior knowledge of whether speech samples were
produced by the same or different speakers. The Cllr acts as an error measure
that captures the gradient goodness of a set of likelihood ratios derived from
test data (Morrison 2009b p. 6). Previous studies of LRs for forensic speaker
comparisons have shown that Cllr proves appropriate for measuring errors
126
(Morrison and Kinoshita, 2008; Morrison, 2011). The equation commonly used
for calculating Cllr is provided in Equation (1).
(1)
from Morrison (2009, p. 2391)

Nss = Number of same speaker pairs
Nds = Number of different speaker pairs
LRss = LR from same speaker pairs
LRds = LR from different speaker pairs
Cllr was calculated using Brmmers FOCAL toolkit23 function cllr.m with the
log-LRs as input. Values of Cllr that are closer to zero indicate that error is low.
For values approaching one the error is considered to indicate poor
performance, while values above one indicate very poor performance (van
Leeuwen and Brmmer, 2007, pp. 343-344).
EER unlike Cllr provides a hard (i.e. binary) accept-reject measure of
validity. This is based on the point at which the percentage of false hits (DS
pairs that ostensibly offer support for the Hp) and the percentage of misses (SS
pairs that appear to offer support for the Hd) are equal (Brmmer and du Preez,
2006, p. 230).
4.6.2 Results
The following sections detail the results of two sets of calculations of the
discriminant potential of AR. The first investigates the capacity of AR to
discriminate within a large homogeneous group of 100 males. The second
23 http://sites.google.com/site/nikobrummer/focal (Downloaded: 13 August 2012)
127
considers how the minimum number of syllables in a speech interval may affect
the strength-of-evidence for AR.
4.6.2.1 LRs for ARs in 100 SSBE Male Speakers
The results for the calculation of LRs on ARs are summarized in Table
4.7. The second row contains the results from SS comparisons and the third row
contains DS comparison results. The total percent of correct SS and DS
comparisons is shown in the second column. Correct LRs are determined by
whether or not an LLR for SS comparisons is a positive value (providing support
for the prosecution hypothesis) and whether an LLR for DS comparisons is a
negative value (providing support for the defense hypothesis). The third
through fifth columns report on the strength-of-evidence, whereby the third
column presents the mean LLR for all comparisons (either SS or DS). The
smallest calculated LLR is in the fourth column, followed by the largest
calculated LLR. The final two columns provide the EER and Cllr values for the
entire system.
Table 4.7: Summary of LR-based discrimination for articulation rate (100 speakers)
Comparisons % Correct Mean LLR Min LLR Max LLR EER Cllr
AR SS 90.0 0.18 -1.48 2.06
.3340 .8981
AR DS 46.2 -2.94 -8.76 0.82
Table 4.7 shows that AR performs much better with SS comparisons than DS
comparisons. The results may seem counterintuitive, since there is higher
within-speaker variability than between-speaker variability for AR, and it might
be assumed that the high within-speaker variation would cause DS pairs to
perform better than SS pairs. However, it appears that, because the degree of
128
variation in AR is so high within speakers overall, the system tends to allocate
higher degrees of similarity if two speakers have similar degrees of (high)
within-speaker variation. This is evident in the fact that for DS comparisons, the
system performs slightly worse than chance (which is 50%, since an LLR
correct/incorrect response is categorized as supporting the Hp or the Hd,
respectively) as the AR system tends to over-predict pairs being from the same
speaker rather than different speakers (note the high error rate in correct DS
judgments) 24. Table 4.7 shows that Cllr is approaching one, but is still under it.
Following van Leeuwen and Brmmer (2007) this would be classified as a
poor performance. The EER is high at 33.4% for AR as a system and the mean
SS LR offers only limited evidence to support the prosecution hypothesis (Hp).
The mean DS LR is slightly stronger, and offers moderate evidence to support
the defense hypothesis (Hd).
The Tippett plot in Figure 4.6 provides a visual measure of the
performance of AR as a discriminant feature. The x-axis displays log10 LRs
where zero is the division between support for Hp (>0) and support for Hd (<0).
The y-axis displays cumulative proportion. Flatter contours indicate a higher
proportion of pairs that achieve a stronger strength-of-evidence, and contours
that are steeper indicate a weaker strength-of-evidence. The results for SS and
DS comparisons are assessed together.
24 An additional explanation for the poor performance of DS comparisons could be that the
system is not optimally calibrated (see 8.5.3). This is evident in the intersection between the
SS and the DS distributions in Figure 4.6, as the intersection is not at LLR = 0, but further to the
right into the higher scores. Therefore, many DS comparisons obtain an LLR larger than zero.
This miscalibration is also potentially the reason for the poor Cllr values.
129
Same speaker comparisons
Different speaker comparisons
Figure 4.6: Tippett plot of articulation rate
Figure 4.6 shows that error rates are higher for DS comparisons than they are
for SS comparisons. The SS line is steeper than that of the DS line and provides a
relatively low strength of evidence. DS, on the other hand, can attain higher
strength of evidence (a Log10 LR of -5 or even lower), although these values are
reserved for a very small percentage of DS comparisons. It is important to
remember when analyzing SS and DS LR results that two samples cannot get
more similar for a feature than identical (Rose et al., 2006, p. 334), and
therefore DS comparisons will always carry the potential for achieving a higher
strength of evidence than SS comparisons. The Tippett plot paints an overall
picture that AR as an individual parameter is relatively weak at discriminating
between individuals, and only produces higher strength of evidence for a very
small proportion of DS comparisons.
130
4.6.2.2 LRs for ARs of 25 Speakers with Variation in the Minimum Number
of Syllables in a Speech Interval
The following section reports on the LRs calculated for the ARs of 25
speakers while varying the minimum number of syllables required in a speech
interval. Table 4.8, which has a similar structure to that in Table 4.7, provides
the results of the different systems. The first column includes a value next to the
SS or DS lines to indicate the minimum number of syllables required for given
speech intervals in that system.
Table 4.8: Summary of LR-based discrimination for mean articulation rate when varying the
minimum number of syllables in a speech interval (25 speakers)
4 < Same Speaker 84.6 0.005 -1.800 0.369
0.2500 1.0260
4 < Different Speaker 52.6 -0.252 -3.153 0.771
5 < Same Speaker 76.9 -0.014 -1.703 0.340
0.3109 1.0326
6 < Same Speaker 69.2 -0.068 -1.73 0.412
0.3846 1.0976
7 < Same Speaker 53.8 -0.152 -1.908 0.435
0.4615 1.1780
8 < Same Speaker 69.2 -0.103 -2.731 0.662
0.3782 1.1104
9 < Same Speaker 61.5 -0.135 -1.243 0.414
0.4615 0.9958
10 < Same Speaker 53.8 0.024 -1.175 0.322
0.5385 0.9848
Increasing the minimum number of syllables for a speech interval in the
different systems did not improve the percentage of correct SS or DS
comparisons; however, it does improve the Cllrs, as seen in the rightmost
column in Table 4.8. By increasing the minimum number of syllables required
for a speech interval, the percentage of correct SS comparisons drops by 30.8%.
For DS comparisons that number drops by 15.4%, with Cllr improving by .0412.
131
EER is more erratic, showing both increases and decreases in its performance
as the minimum syllable number in an interval is increased (although it is
mainly an upward trend).
In comparison to the system discussed in 4.6.2.1 for 100 speakers, the
percentage of correct SS comparisons and the Cllr value are worse, while the
percentage of correct DS comparisons is slightly improved. Overall, EERs for
minimum syllable lengths above five perform worse than the AR system in
4.6.2.1. Caution should be exercised when interpreting these results, however,
as there were only 13 test speakers and 12 reference speakers, whereas the
tests described in 4.6.2.1 included 50 speakers in both test and reference sets.
4.6.3 Discussion
Although within-speaker variation can be decreased by increasing the
minimum number of syllables in a speech sample, it appears that the system
performs worse overall in terms of the percentage of speaker comparisons
judged correctly, in spite of the Cllr improving slightly. However, no system
produced a Cllr better than that seen in 4.6.2 with all 100 speakers and a
minimum of four syllables per speech interval. It is important to note that, by
increasing the minimum number of syllables required in a speech interval for
the 25 speakers, the number of useable speech intervals available for each
individual is decreased, potentially affecting the results. This means that
systems were potentially performing worse due to a decrease in the number of
speech intervals available; or rather, the change is the result of a combination of
this along with the increase in the minimum syllables required for a speech
interval.
132
Most importantly, by re-redefining the minimum syllable count
requirement for a speech interval the overall performance of a system is
susceptible to changes in data collection and methodologies for AR. The need
for consistency in analysis techniques for all forensic domains, including AR in
forensic speech science, may be best achieved through prescribed methods
(based on rigorous empirical testing) for calculation.
4.7 Conclusion
Overall, it appears that AR can be classified as a speech parameter that carries
higher intra-speaker variation than it does inter-speaker variation. In respect of
the large number of available methodologies for the calculation of AR, it appears
that the defining of a speech interval (memory stretch vs. inter-pause unit) does
not have a significant effect on the results. However, in the context of real
forensic casework if methodologies and analysts are to be kept consistent for
the analysis of suspect and criminal samples then problems relating to AR
calculation methods will be minimized.
AR as a discriminant parameter has proved to be a very poor one25, and
it is not anywhere close to being as good at discriminating between individuals
as experts have claimed it to be (Chapter 3). This raises the question as to why
some analysts are using it at all in casework except for instances of very high or
low AR. However, exceptions exist for those speakers that are classified as
outliers. It has been shown that AR offers a very weak strength of evidence for
SS comparisons, while DS comparisons can potentially offer a higher level of
strength of evidence. However, this must be traded off against the fact that they
25 At least as these results suggest.
133
produce a very high rate of incorrect DS judgments. Despite all efforts to
decrease within-speaker variation by increasing the minimum number of
syllables in a speech interval, the overall system is not improved from the
original system results shown in 4.6.2. The Cllr in the best-performing system
is .8981, which classifies it as performing poorly since the score is close to 1,
following Brmmer and du Preez (2006).
The analysis of AR as a parameter under an LR framework in forensic
speech science urges caution for casework, in that parameters previously
thought to be good speaker discriminants might transpire to carry higher intra-
speaker variation than inter-speaker variation, which will generally result in a
lower strength of evidence for a given parameter. More research on speaker
discriminants for other commonly-used parameters in forensic casework is
clearly needed, because there is a risk that some experts in the field are
analyzing certain features rather blindly. That is to say, they are giving weight
to features which actually provide little in terms of discrimination power. This
is shown by the fact that 93% of experts surveyed analyze speech tempo, and
73% of those do with varying regularity; furthermore, 20% of experts reported
speech tempo to be the single most useful discriminant. The analysis carried out
in this chapter provides evidence that AR may be a far from useful discriminator
in many cases.
Although AR may not be the discriminant shibboleth all experts hope for,
it is important that AR is still considered in forensic speaker comparisons in
conjunction with other speech parameters. There are instances in which
speakers may have a very low or high AR, and in which the parameter can be
considered useful (either as evidence for or against speaker identity). As Rose
134
(2007a, p. 1820) points out not all speakers differ from each other in the same
way. Therefore there will be a few individuals for whom AR is potentially a
good discriminant parameter, as was evident in the small number of high LRs
for DS comparisons in 4.6.
135
Chapter 5 Long-Term Formant
Distributions
5.1 Introduction
Forensic speech science literature is largely characterized by research
investigating the discriminant power of vowels. In Chapter 3, it was reported
that 28% of forensic phoneticians found vowels to be the most useful feature in
discriminating between speakers; overall, this places vowels as the second
highest-ranked discriminant parameter (along with accent/dialect variants)
among all possible parameters analyzed for FSCs. Long-term formant
(frequency) distribution (LTFD) work to date constitutes only a small portion of
the research carried out. LTFD is the method used to calculate the average
values for each formant of a speaker over a given speech recording. For a given
formant (i.e. F1-F4), measurements for all vowels produced by a single speaker
are averaged across the entire recording or relevant portions of the recording.
This means that for each formant (F1-F4) of a speaker there is an LTFD value
and a standard deviation (SD), which will be referred to as LTFD1, LTFD2,
LTFD3, and LTFD4. LTFDs are frame-by-frame measurements (5 msecs in
length for the current study); therefore, long vowels carry more weight than
short vowels in that they yield a greater number of measurements per vowel. A
positive attribute of LTFDs is that they do not require the categorization of
individual vowels into phoneme classes, as all vowels are considered in an
analysis. This results in greater time savings. LTFD also avoids the potential
correlations between vowel phonemes which would not allow those correlated
136
phonemes to be combined, since combinations of parameters under Bayes
theorem are only allowed if events can be shown to be mutually exclusive.
This chapter provides population statistics for LTFD while also
investigating the discriminant power of LTFD under a likelihood ratio
framework. Population statistics which consist of LTFD means and SDs are
reported for F1-F4, in order to give indications of between- and within-speaker
variation. LRs are calculated for LTFD1-4 individually as well as in different
combinations relevant for casework. The results of the LRs are presented and
considered in terms of strength of evidence, Cllr, EER, and proportion of SS and
DS comparisons that were correctly identified.
The results of vowel research in forensic phonetics are well documented
in the literature as a way in which to characterize the speech of an individual.
Many different methods have been offered for acoustic analysis, the most
common being temporal mid-point center-frequency measurements of
formants for different vowels (Jessen, 2008; Rose, 2002; 2006a, b, c; Rose et al.,
2003). Investigations have also been carried out using formant dynamics in
order to capture the trajectories of specific vowels (McDougall, 2004;
McDougall and Nolan, 2007; Hughes, 2011; Hughes et al., 2009). Formant
dynamic research revealed that formants appear to be consistent within
speakers and that there is variation with respect to formants between speakers.
This research led to the argument that the development of techniques for
measuring dynamic features should be given more attention (McDougall, 2006).
Long-term spectra (LTS) were developed with this in mind, as a means of
137
capturing the average of all spectral slices in a recording. However, LTS
considers voiced speech as well as voiceless portions which could potentially
include background noise (Nolan and Grigoras, 2005). LTFD, like LTS, was also
developed to identify the dynamism of formants, but it is only concerned with
the vowels and certain voiced portions in a recording (Nolan and Grigoras,
2005). There have also been a number of analyses concerned with vowels
under a LR framework (e.g. Alderman, 2004; Kinoshita and Osanai, 2006; Rose,
2007a; Morrison, 2009). However, only three have considered LTFD (Becker et
al., 2008; French et al., 2012; Jessen et al., 2013).
Nolan and Grigoras (2005) were the first to report the use of LTFD for
forensic speaker comparisons. In their study, the authors consider analysis of
LTFD1 and LTFD2 in order to eliminate a suspect who is thought to have made
some obscene phone calls. The first author carried out an auditory analysis of
vowels and took mid-point center-frequency measurements of monophthongs.
Diphthongs were also included in the analysis and the beginning and end points
of the vowels were measured. In general, the vowel analysis by the first author
suggested that the speech in the criminal samples and the speech in the suspect
sample were poorly matched. Each vowel in the suspect samples exhibited
systematic differences from those found in the criminal samples, thereby
rendering the criminal and suspect samples incompatible. Given that there were
other equally valid methods to arrive at acoustic characterization of speakers,
the second author carried out a re-analysis using alternative approaches.
The second author had previously developed new techniques for
speaker comparison (Grigoras, 2001; 2003) that included speaking
fundamental frequency (SFF), LTS, and LTFD. All three approaches were carried
138
out using Catalina Toolbox26. LTFD was used in the re-analysis of the data in the
obscene phone call case. This approach considers the long-term disposition of
formants (Nolan and Grigoras, 2005, p. 162), an aspect which LTS fails to
grasp. Only the voiced frames in the recordings were used for analysis and
linear prediction was used to estimate LTFD1-4. The results from LTFD1-4 give
an overall indication of each formant distribution, from which it is clear that
LTFD2 and LTFD3 in the criminal recordings are considerably lower than in the
suspect recording. LTFD4 is also shown to be relatively higher in the criminal
samples than the suspect sample. Given the distribution of LTFD, this analysis
gave further substance to the argument that there were two different speakers
involved and that the suspect and criminal were not one and the same person.
With respect to the approaches employed in the re-analysis, LTFD provided a
very clear picture of the average behaviour of each formant (Nolan and
Grigoras, 2005, p. 169). It also provides strong insights into the dimensions of a
speakers vocal tract where these are reflected in the maximum LTFD27. The
formant frequency values for LTFD are inversely related to the speakers vocal
tract size, whereby a longer vocal tract will result in lower formant values
(Nolan and Grigoras, 2005; French et al., 2012). LTFD also has the capacity to
indicate certain habits speakers use, such as palatalization, which are indicated
by a raised LTFD2 (Nolan and Grigoras, 2005). French et al. (2012) also show
that voice qualities related to tongue body position are correlated with LTFD.
Additionally, the shape of the distributions for the estimates of each formant is
useful in identifying speakers who have either more or less variable formants.
26Available at http://www.forensicav.ro/download.htm
27The maximum LTFD (for LTFD1 and LTFD2) is reflected in the overall area of a speakers
vowel space.
139
This is classified by distributions which are leptokurtic (narrow-peaked) or
platykurtic (broad-peaked). Although LTFD provides promising
characterizations of individuals speech it fails to reveal the variation that exists
in specific vowel segments.
Moos (2010) utilized the LTFD method detailed in Nolan and Grigoras
(2005) and analyzed LTFD2 and LTFD3 values of mobile phone speech in both
read and spontaneous speech of 71 male German speakers from the Pool 2010
corpus (Jessen et al., 2005). The spontaneous speech was elicited from speakers
while having them describe objects to another person (their compatriot)
without using certain proscribed words, similar to the strategies in the board
game Taboo. The person who was matched with the speaker feigned
ignorance in the exchange in order to encourage more thorough descriptions
and longer stretches of speech. The read speech was produced by speakers
reading a German version of The North Wind and the Sun (Moos 2010).
Recordings were edited to include only the vocalic portions, where laterals,
approximants, vocalic hesitations, and creaky voice were part of that stream.
Nasals, areas of strong nasality, and vowels spoken on a high pitch (where
individual harmonics were visible in the spectrogram rather than formants)
were not included. After cutting down the recordings in Wavesurfer, the length
of the spontaneous speech was between 12 and 83 seconds (mean = 40 sec) and
the length of the read speech was between 8 and 16 seconds (mean = 12 sec).
Moos (2010) found LTFD3 values to be slightly more helpful28 than
LTFD2 in terms of speaker characterization because, overall, the former had
28This is also seen in Simpson (2008) and Clermont et al. (2008) with regard to F3 values for a
number of phonemes that were measured using mid-point center frequencies.
140
smaller intra-speaker variation. The LTFD values from read speech were
reported as being higher than those in spontaneous speech. However, this could
be due in part to the fact that the The North Wind and the Sun is a tool used in
phonetic experiments to get an array of phonemes and tokens. For this reason,
the vowel spaces of speakers could be artificially enlarged, returning a greater
spread of LTFD values per speaker. It is important to note that details of the
spontaneous speech that was elicited were not provided, and could have in
theory given a similar distribution of phonemes and tokens. However, that is
unlikely, as spontaneous speech tends not to provide as wide of an array of
phonemes as is usually the case with read speech (of course this is dependent
on the chosen text). It was also noted by Moos (2010) that it is vital to know
whether there is a sufficient amount of data in order to analyze LTFD; 6 seconds
of pure vocalic stream were suggested as a minimum. Overall, Moos (2010)
classifies LTFD as a valuable measure to include in forensic speaker
comparisons and identifications. LTFD was also found to be independent of (i.e.
not correlated to) F0, dialect, and speech rate, making LTFD viable for
combination with these parameters under an LR conclusion framework.
Becker et al. (2008) investigated the use of Gaussian Mixture Models for
LTFD1-3 under an LR framework (see Jessen et al., 2013 for a similar LTFD
study but using the software Vocalise29). Spontaneous speech was used from 68
male German speakers from the Pool 2010 corpus recorded in a laboratory
setting. The speech had been elicited as described above in Moos (2010), and as
in Moos (2010), the data used in Becker et al. (2008) were transmitted through
mobile phone connections in order to simulate forensically-relevant recordings.
29 http://www.oxfordwave research.com/j2/products/vocalize [ Accessed: 8 August 2013]
141
These recordings were then edited to remove consonantal information as well
as portions of speech where formant structures were not clearly visible.
Formant tracking was then used to identify peaks and LTFD measurements
were extracted in Wavesurfer. The formant measurements for the first half of
each recording were used as a training set, and the test set consisted of the
second half of the formant measurements. Those formant measurements in the
test set were halved again in order to increase the possible number of
comparisons. This resulted in recordings from the training set being around 22
seconds in length, while those in the test set were around 11 seconds long.
LTFD1-3 as well as their corresponding bandwidths were considered in the
analysis. 18 speakers measurements were used to create the Universal
Background Model (UBM), and one Gaussian Mixture Model (GMM) was
estimated for the reference population, using 8 mixtures. The remaining 50
speakers were used in the test and a total of 100 same-speaker comparisons
and 4,900 different speaker comparisons were carried out.
The lowest (i.e. the best performing) EERs were found for combinations
that included bandwidths (BW), these being
LTFD1+LTFD2+LTFD3+BW1+BW2+BW3 and LTFD1+LTFD2+BW1+BW2,
which achieved EERs of 0.030 and 0.042, respectively. The lowest EER in which
BW was not included was the combination of LTFD1+LTFD2+LTFD3, which had
an EER of 0.053. Overall, discrimination levels were high, and Becker et al.
(2008) note that the speaker models created using LTFD can relate directly to
the configuration of the vocal tract, in turn perhaps revealing speaker-specific
variations in the distributions.
142
Jessen and Becker (2010) build on Becker et al. (2008) and investigate
LTFD as a speaker discriminant further. They first consider the relationship
between LTFD2-3 and body height for 81 male speakers from the Pool 2010
corpus of telephone transmitted speech. Both LTFD2 and LTFD3 were found to
have significant negative correlations with stature, where taller individuals
were associated with lower LTFD2 and LTFD3. The Pearson correlation
coefficients were both just over 30% (r = -0.316 and -0.339 respectively), but
were nonetheless significant at the 1% level. Jessen and Becker (2010) then
examined the consistency with which analysts measure LTFD. Five phoneticians
measured LTFD2 and LTFD3 for 20 speakers from the Digs dialect corpus
(Jessen and Becker, 2010). LTFD means were compared across analysts and it
was found that LTFD2 had Pearson correlations between 0.84 and 0.95, while
for LTFD3 these figures were between 0.98 and 0.9930. Consistency across
analysts was higher for LTFD3 than for LTFD2, but both formants achieved
highly consistent results overall, showing that the methodology for LTFD
analysis is easily replicable. It has been posited that LTFD is potentially
language-independent, as all vowel phonemes are averaged (Nolan and
Grigoras, 2005). Jessen and Becker (2010) tested this hypothesis using three
speakers of different German dialects in the Digs dialect corpus, as well as
Russian and Albanian speakers under analogous recording conditions. They
found that the different languages did not appear to differ in terms of the LTFD-
space that they occupy (one-way ANOVA [F(4,55) = 0.44; p = 0.77]).
30 It is unclear if these figures refer to r or r2 values.
143
The authors then investigated the effects of Lombard speech31 on LTFD.
Mean LTFDs from 31 speakers in the Pool 2010 corpus (telephone-transmitted
speech) were used to analyze possible Lombard effects on LTFD1-3. LTFD1 was
found to be consistently higher in the Lombard condition than in the modal
condition, with high levels of intra-speaker variation present. LTFD2 and LTFD3
were inconsistent in their effects across speakers, and both yielded non-
significant differences. Although LTFD1 was shown to be affected by Lombard
speech, it is often of limited use in forensic casework, due to the effect of
telephone transmission on F1 (Byrne and Foulkes, 2004; Jessen and Becker,
2010; Knzel, 2001). Finally, the authors tested the performance of LTFD
analysis modeled using GMMs against ASR. They found ASR to outperform LTFD
analysis, with an EER of 0.107 for ASR as compared to an EER of 0.243 for LTFD.
Logistic-regression fusion (which accounts for correlations between resulting
LRs and then applies statistical weightings) was also used to try to improve
results (EER= 0.108). However, it still performed worse than ASR on its own.
The results found by Jessen and Becker (2010) were promising for LTFD use in
forensic casework. LTFD revealed a negative correlation of individual formants
with body height, a high consistency in measurements across different
phoneticians, the potential to use LTFD statistics from one language across
many languages, and the limited effect of Lombard speech on LTFD2 and
LTFD3.
Most recently, French et al. (2012) examined LTFD in conjunction with
Mel-Frequency Cepstral Coefficients (MFCCs; abstract properties of the acoustic
31Lombard speech is the tendency of speakers to increase their vocal effort when speaking
(typically due to loud noise; Lombard (1911)).
144
signal, which are reflections of the dimensions of the vocal tract) and voice
quality (VQ). They considered the efficacy and limitations of the three
parameters, and correlations among the parameters. The study used the
recordings from Task 2 of the DyViS database (Nolan et al., 2009). All original
recordings were 15 to 25 minutes in duration. The recordings were edited to a
minimum of 50 seconds of vowels per speaker32. The iCAbS (iterative cepstral
analysis by synthesis) formant tracker (Clermont et al., 2007) was used to
automatically extract and measure F1-F4 every 5 msec. This yielded between
10,000 and 30,000 F1-F4 measurements per speaker. LRs for LTFD1-4 were
calculated using UBM-GMM in the same way as that detailed in Becker et al.
(2008). In total there were 200 same speaker (SS) comparisons and 9800
different-speaker (DS) comparisons, as each recording for a speaker was
divided in half. French et al. (2012) found LTFD1-4 to perform very well, with
97.4% of DS comparisons and 94% of SS comparisons identified correctly. In
comparison with the discrimination levels of the MFCCs and VQ on the same
data set, LTFD achieved similar error rates to the other methods; one method
did not significantly outperform another. In terms of correlations between
LTFD, MFCCs, and VQ, there were correlations found between the LRs produced
from MFCCs and LRs calculated from the UBM-GMM analysis of LTFDs (r =
0.39). There was a weak correlation identified between VQ and LTFD globally (r
= 0.12). However, there were some specific aspects of VQ that were more
closely correlated with single LTFD measurements (e.g. raised larynx and
LTFD1, r = 0.40). In conclusion, French et al. (2012) suggest the use of a vocal
32This was done using Synthesis Toolkit CV software that was adapted by Philip Harrison from J
P French Associates.
145
tract output measurement (e.g. MFCC, VQ, LTFD) to be used as one of many
tools in forensic speaker comparisons in order to examine speech as varying
human behavior.
The research in the studies presented above was carried out using a
traditional (acoustic) phonetic approach, or an MFCC-based one. The
discriminant performance of LTFD was tested in previous studies only using the
GMM-UBM LR framework. At present, no previous studies have provided
population statistics for LTFD in English or considered LTFD under an MVKD LR
framework. This chapter will address both gaps.
5.3 Population Statistics for LTFD1-4
The following section discusses the collection of population statistics for LTFD
in a large, linguistically-homogeneous group of 100 male speakers. These data
serve as the first of their kind in providing detailed information on the
distribution and variation that occurs in LTFD for a large group of individuals
who speak Southern Standard British English (SSBE).
5.3.1 Methodology
Spontaneous speech recordings of 100 male speakers of SSBE, aged 18-
25, were analyzed. The recordings were from Task 2 (a conversation between
the speaker and his accomplice) of the DyViS database (Nolan et al., 2009). The
recordings were automatically segmented to obtain a minimum of 50 seconds of
concatenated vowels per speaker, and the iCAbS formant tracker (Clermont et
al., 2007) was used to automatically extract and measure F1-F4 every 5 msec.
146
Population statistics were calculated by averaging all measurements for each
formant for a single speaker and also taking the SD for each of those formants.
5.3.2 Results
The following section analyzes LTFD1-4 individually. The distributions
for LTFD1 means and SDs are provided in Figures 5.1 and 5.2. The y-axis
represents the number of speakers with mean LTFD formant frequencies that
fall within a given range and the x-axis represents 10Hz-wide formant
frequency bins, presented in Hertz (Hz).
Mean LTFD1
18
Number of Speakers
16
14
12
10
8
6
4
2
0
Formant Frequencies (Hz)
Figure 5.1: Distribution of mean LTFD1
Figure 5.1 shows a normal distribution with a slight negative skew due to
suspected outliers ( 1.5 times the interquartile range) at 364.7Hz, 367.1Hz,
375.6Hz, 386.7Hz and one suspected outlier at 515.6Hz. The overall mean for
the group LTFD1 is 451Hz, with a range of 364.7Hz to 515.6Hz. The SD of the
means is 29.9Hz, and all 100 speakers SDs fall within a 150.9Hz range.
147
Standard Deviations for LTFD1
18
Number of Speakers
16
14
12
10
8
6
4
2
0
Standard Deviation (Hz)
Figure 5.2: Distribution of standard deviation in LTFD1
The standard deviation values for LTFD1 within speakers follows a roughly
normal distribution. There are three suspected outliers at 201.6Hz, 203.6Hz,
and 209.8Hz. The mean SD is 131.4Hz, with a range of 64.8Hz to 209.8Hz. The
SD of the mean SDs is 26.8Hz. All 100 speakers have SDs within 145Hz, which is
a larger range (by 58.7Hz) than the range of means found in Figure 5.1.
The cumulative distribution graphs of LTFD1 means and SDs in Figure
5.3 and Figure 5.4, respectively, show the percentile within which a given
LTFD1 mean or SD falls within the population. The y-axis is the cumulative
proportion, and the x-axis represents formant frequencies in Hz.
148
Mean LTFD1
100

90
80
70
60
50
40
30
20
10
0
355 365 375 385 395 405 415 425 435 445 455 465 475 485 495 505 515 525
Figure 5.3: Cumulative percentages for mean LTFD1
The curve in Figure 5.3 is characterized by steepness in the central portion and
gentle gradients in the first and third portions. Despite the steepness of the
central section, the curve is overall a lot more gradient than was seen for AR in
Chapter 4. 1 SD from the mean gives a range between 421.1Hz and 480.9Hz,
into which roughly 83% of the population tested here falls. The cumulative
distribution of individual SDs is illustrated in Figure 5.4.
149
100

90
80
70
60
50
40
30
20
10
0
55 70 85 100 115 130 145 160 175 190 205
Figure 5.4: Cumulative percentages for standard deviation of LTFD1
The curve in Figure 5.4 is slightly steeper than that seen in Figure 5.3, and the
beginning and end portions are less gradient. 1 SD from the mean SD gives a
53.6Hz range between 104.6Hz and 158.2Hz, into which roughly 77% of the
population falls. Given that the mean LTFD values are representative of the
variation that occurs between speakers, and the SD values represent within-
speaker variation, a variance ratio (Rose et al. 2006) can be calculated to
ascertain which variation is higher. The LTFD1 variance ratio is 0.05, which is
indicative of higher inter-speaker variation than intra-speaker variation.
The results for LTFD2 are presented in Figures 5.5 and 5.6. The graphs
illustrate the population distributions for LTFD2 mean and SD.
150
Mean LTFD2
20
Number of Speakers
18
16
14
12
10
8
6
4
2
0
Figure 5.5 has as a normal distribution with a slight positive skew due in part to
a suspected outlier at 1633Hz. The overall mean LTFD2 for the group is
1476.7Hz, with a range of 1363.9Hz to 1633Hz. The SD of the LTFD2 means is
55.9Hz, and all 100 speakers fall within a 269.1Hz window.

18
Number of Speakers
16
14
12
10
8
6
4
2
0
240-250
250-260
260-270
270-280
280-290
290-300
300-310
310-320
320-330
330-340
340-350
350-360
360-370
370-380
380-390
390-400
400-410
410-420
420-430
430-440
151
The standard deviation for LTFD2 within speakers is again roughly normally
distributed. There are two suspected outliers at 425.4Hz and 437.7Hz. The
LTFD2 mean SD is 322.7Hz, with a range of 249.3Hz to 437.7Hz. The SD of the
mean SDs is 37.7Hz. All 100 speakers have SDs within 188.4Hz.
The cumulative distribution graphs of means and SDs in Figure 5.7 and
Figure 5.8, respectively, show the percentiles at which a given LTFD2 mean or
SD falls within the population.
Mean LTFD2
100
90
80
70
60
50
40
30
20
10
0
1350 1380 1410 1440 1470 1500 1530 1560 1590 1620 1650
The curve in Figure 5.7 is rather gradual compared to those in Figures 5.3 and
5.4, which is indicative of the relatively platykurtic distribution seen in Figure
5.5. 1 SD from the mean gives a range between 1420.8Hz and 1532.6Hz, into
which roughly 72% of the sample population falls. The cumulative distribution
of individual LTFD2 SDs is illustrated in Figure 5.8.
152
100

90
80
70
60
50
40
30
20
10
0
240 255 270 285 300 315 330 345 360 375 390 405 420 435
Figure 5.8: Cumulative percentages for standard deviation in LTFD2
The curve in Figure 5.8 is similar to that in Figure 5.7; however, the rate of
increase of the middle portion in Figure 5.8 is much more variable. 1 SD from
the mean SD gives a range between 285Hz and 360.4Hz, into which roughly
71% of the sample population falls. Comparing the intra-speaker variation to
the inter-speaker variation for LTFD2, there is a variance ratio 0.03. This
indicates higher levels of variation within speakers than between them.
The results for LTFD3 are presented in Figures 5.9 and 5.10 below. The
graphs represent the population distributions for LTFD3 mean and SD,
respectively.
153
Mean LTFD3
25
Number of Speakers
20
15
10
Figure 5.9 has as a normal distribution with a slight negative skew. This is
potentially due in part to a suspected outlier at 2824.4Hz. The overall mean
LTFD3 for the group is 2478.5Hz, with a range of 2212.6Hz to 2824.4Hz. The SD
of the LTFD3 means is 106.5Hz, and all 100 speakers fall within a 611.8Hz
window.
154
18
Number of Speakers
16
14
12
10
8
6
4
2
0
Figure 5.10: Distribution of standard deviation distribution in LTFD3
The standard deviation for LTFD3 within speakers is roughly normally
distributed with a slight negative skew. It can be expected that the skew is due
to the extreme outlier at 516Hz and three suspected outliers at 416.3Hz,
422.1Hz, and 491.9Hz. The mean SD is 277.9Hz, with a range of 168.2Hz to
516Hz. The SD of the LTFD3 SDs is 62.8Hz. All 100 speakers have SDs within
188.4Hz of each other.
5.11 and Figure 5.12, respectively, illustrate the percentile at which a given
LTFD3 mean or SD falls within the population.
155
Mean LTFD3
100

90
80
70
60
50
40
30
20
10
0
2175 2250 2325 2400 2475 2550 2625 2700 2775 2850
The curve in Figure 5.11 is rather steep in the middle portion, with the
beginning and end portions of the slope being more gradient. 1 SD from the
mean gives a range between 2372Hz and 2585Hz, into which roughly 70% of
the population falls. The cumulative distribution of individual speakers SDs for
LTFD3 is presented in Figure 5.12.
156
100

90
80
70
60
50
40
30
20
10
0
150 175 200 225 250 275 300 325 350 375 400 425 450 475 500 525
The curve in Figure 5.11 is rather steep, with the end portion of the slope, which
starts around 375Hz, being more gradient than the beginning. 1 SD from the
mean gives a range between 215.1Hz and 340.7Hz, into which roughly 68% of
the sample population falls. The calculated variance ratio for LTFD3 is 0.15. This
suggests it is more likely that one will find higher levels of variation within a
speaker than between that speaker and the rest of the population, which was
also the case for LTFD1 and LTFD2.
The results for LTFD4 are presented in Figures 5.13 and 5.14 below. The
graphs display the population distributions for LTFD4 mean and SD,
respectively.
157
Mean LTFD4
16
Number of Speakers
14
12
10
8
6
4
2
0
Figure 5.13 shows that mean LTFD4 has as a fairly normal distribution, though
with a slight negative skew. However, there are no suspected (1.5 x the
interquartile range) or definite (3 x the interquartile range) outliers in the data.
This could simply be the natural distribution of the data or perhaps it is
revealing measurement errors that occurred for those individuals that appear
to have lower LTFD4 means. The overall mean LTFD4 for the group is
3660.9Hz, with a range of 3249.9Hz to 4019.5Hz. The SD of the LTFD4 means is
170.9Hz, and all 100 speakers fall within a 769.6Hz window.
158
18
Number of Speakers
16
14
12
10
8
6
4
2
0
The standard deviation for LTFD4 within speakers is roughly normally
distributed. There are three suspected outliers at 633.3Hz, 640.5Hz, and
649.5Hz. The mean SD is 482.2Hz, with a range of 356.8Hz to 649.5Hz. The SD
of the LTFD4 SDs is 67.2Hz. All 100 SSBE speakers have SDs within 292.7Hz of
each other.
5.15 and Figure 5.16, respectively, show the percentile at which a given LTFD4
mean or SD falls within the population.
159
Mean LTFD4
100

90
80
70
60
50
40
30
20
10
0
3200 3300 3400 3500 3600 3700 3800 3900 4000 4100
The curve in Figure 5.15 is characterized by a rather gradual increase, which is
more similar to the curve in Figure 5.7 than it is to the curves illustrated in
Figures 5.3 and 5.11. 1 SD from the mean gives a range between 3490Hz and
3831.8Hz, in which roughly 63% of the sample population falls. The cumulative
distribution of individual SDs for LTFD4 is illustrated in Figure 5.16.
160
100

90
80
70
60
50
40
30
20
10
0
340 370 400 430 460 490 520 550 580 610 640
The curve in Figure 5.16 has a rather gradual, almost linear slope. The graph
has a similar shape to the curve shown in Figure 5.15. 1 SD from the mean SD
gives a range between 415Hz and 549.4Hz, into which roughly 65% of the
population falls. The variance ratio for LTFD4 is 0.13, which again indicates that
one will be more likely to find higher levels of variation within a speaker than
between that speaker and the rest of the population, which we also saw in the
case of LTFD1-3. Given that variance ratios greater than one indicate that more
variation occurs within individuals than between them, the LTFDs with higher
variance ratios discriminate better between individuals than do those with
lower variance ratios. LTFD3 had the highest ratio at 0.15, followed by LTFD4 at
0.13, LTFD1 at 0.05, and finally LTFD2 at 0.03. The LR results in the next section
( 5.4) are hypothesized to follow these predictions.
161
5.3.3 LTFD1-4 Results Compiled
For simplification purposes the overall LTFD population results are
displayed in Table 5.1 and 5.2. Table 5.1 compiles LTFD1-4 results for between-
speaker variation, while Table 5.2 presents the compilation of LTFD1-4 results
for within-speaker variation. The first columns identify the formant, and the
second through fourth columns contain mean SD, range (of SD), and SD (of SD).
Table 5.1: Overall between-speaker results for LTFD1-4
LTFD Mean (Hz) Range (Hz) SD (Hz)

1 451.0 364.7-515.6 29.9
2 1476.7 1369.9-1633.0 55.9
3 2478.5 2212.6-2824.4 106.5
4 3660.9 3249.9-4019.5 170.9
Table 5.2: Overall within-speaker results for LTFD1-4
LTFD Mean SD (Hz) Range of SD (Hz) SD of SD(Hz)

1 131.4 64.8-209.8 26.8
2 322.7 249.3-437.7 37.7
3 277.9 168.2-516.0 62.8
4 482.2 356.8-649.5 67.2
The mean formant measurements for LTFD1-4 in Table 5.1 are very similar to
those of , where F1 is about 500 Hz, F2 is 1500 Hz, F3 is 2500 Hz, and F4 is
3500 Hz (Johnson, 2003). Given that LTFD is an average across all vowel
phonemes, some type of central (with respect to the vowel space) vowel would
be expected. The results in Table 5.1 and 5.2 also show that LTFD3 and LTFD4
have the smallest ratios of mean SD to mean formant value (277.9: 2478.5 and
482.2: 3660.9). This is indicative of the two higher formants being more stable
within speakers, suggesting that they will be better speaker discriminants than
the lower formants (this also coincides with the variance ratios from 5.3.2
above).
162
5.4 Likelihood Ratios for LTFD
The discriminant results of LTFD are presented in the following section. In the
first part, LR results are provided for individual formants as well as in
combination with other formants. The second part considers the effects that
package length (Moos, 2010) has on LR results for LTFD.
5.4.1 Methodology
Likelihood ratios (LRs) were computed using a MatLab implementation
of Aitken and Lucys (2004) Multivariate Kernel-Density formula (Morrison,
2007) for the 100 male speakers in DyViS Task 2. The MVKD formula was
originally developed for use with evidence that included repeated measures of a
given parameter (Aitken and Lucy, 2004). However, LTFD considers evidence
from all possible vowel categories, resulting in raw data that can be extremely
varied. For this reason, the raw formant data were averaged over 0.5 sec
windows (a total of 100 raw data measurements per formant constituted a
single token) for F1-F4 in order to obtain what Moos refers to as packages
(Moos, 2010). There were 100 to 284 (LTFD1-4) measurements per speaker,
with a mode of 100 tokens. An intrinsic discrimination method was used to
calculate LRs, whereby speakers 1-50 acted as the test set and speakers 51-100
acted as the reference set. LRs are calculated for LTFD1-4 individually as well as
in combinations relevant to forensic casework.
These combinations were chosen with respect to common practices in
the field ( 3.9.1.1). Traditionally, the two formants most commonly used in
casework and sociolinguistic studies are F1 and F2, which are measured in
order to reveal aspects of an individuals vowel space (Ash 1988; Milroy and
163
Gordon, 2003). Some FSS experts still analyze only these two formants (
3.9.1.1), and therefore LTFD1 and LTFD2 are considered in combination.
LTFD research has reported that formants are often prone to variation. It
is common for a case to involve the comparison of material from a telephone
recording (cellular phone or landline) against a directly-recorded sample (often
a police interview). Due to the limited bandwidth of transmission over the
telephone (the 340-3700Hz band) there are many acoustic properties of the
signal that are often affected (Foulkes and French, 2012). The most notable are
an artificial increase in F1 values and formants close to 3700Hz disappearing.
Often F4 is missing from the signal altogether (Knzel, 2001; Byrne and
Foulkes, 2004). A similar effect has also been reported for recordings made
using the video and voice recorders in cellular phones (Gold, 2009). For this
reason, some experts avoid F1 and F4 altogether, meaning that only LTFD2 and
LTFD3 are analyzed. In addition, it is important to note that analysts must be
aware that the distance between the microphone (of the recording device) and
the talker (in conjunction with the room acoustics) can also have effects on
formant measurements (Vermeulen, 2009).
A majority of experts (63%) reported measuring F1-F3 in casework (
3.9.1.1), and therefore LTFD1-LTFD3 are considered, as they are the most
commonly-analyzed combination of formants. Finally, LTFD1-4 are considered
in combination to represent the ideal case where F1-F4 are all measureable.
This also provides the upper boundary in terms of the maximum number of
features within a parameter that can be used (for the given data) to achieve the
best possible performance. All LR results are considered in terms of system
164
performance (EER and Cllr) and the magnitude of strength of evidence
(Champod and Evett, 2000).
5.4.2 Results for LTFD1-4: Individually and in Combination
The results for the calculation of LRs on LTFD1-4 individually are
summarized in Table 5.3. The leftmost column represents the LTFD that was
analyzed and whether it relates to same-speaker (SS) or different-speaker (DS)
comparisons. The second column indicates the percentage of SS or DS
comparisons that were correctly classified. Correct SS comparisons have a log
likelihood ratio (LLR) above zero, and correct DS comparisons have an LLR of
less than zero. The mean LLR is in the third column, followed by the minimum
and maximum LLR in the next two columns. The final two columns present the
EER and Cllr values.
Table 5.3: Summary of LR-based discrimination for LTFD1-4 (100 speakers)
LTFD1 SS 72.0 0.224 -2.158 1.902
.2806 .8840
LTFD1 DS 71.7 -4.858 -68.768 1.993
LTFD2 SS 70.0 0.162 -1.077 1.259
.3165 .8119
LTFD2 DS 67.5 -1.939 -27.814 1.602
LTFD3 SS 88.0 0.288 -8.373 3.743
.1700 1.0731
LTFD3 DS 80.6 -11.857 -139.273 1.734
LTFD4 SS 68.0 0.238 -2.258 1.378
.2214 .8085
LTFD4 DS 80.2 -11.574 -124.808 1.301
Table 5.3 shows that, overall, LTFD3 has the lowest EER (.1700) but the highest
Cllr (1.0731). LTFD4 performed second-best in terms of EER (.2214) and best
for Cllr (.8085). The highest EER was for LTFD2 at .3165, but it had the second-
lowest Cllr (.8119). Overall, SS comparisons achieved a higher proportion of
correct results than did DS comparisons, with the exception of LTFD4, where DS
165
comparisons performed 12.2% better. The strength of evidence (i.e. mean LLR)
is stronger for DS comparisons than SS comparisons; mean LLRs are between
-11.857 and -1.939. For SS comparisons, the magnitude of the strength of
evidence is lower, ranging from 0.162 to 0.288. With respect to Champod and
Evetts verbal scale (2000 p. 240) the LLR scores for SS would not even
constitute limited evidence in support of the prosecution hypothesis. However,
there are some cases where a formant individually achieves a stronger strength
of evidence, as per LTFD3 with its maximum LLR of 3.743 (moderately strong
evidence).
Examining LTFD1-4 individually, the results in Table 5.3 suggest that
LTFD3 performs the best overall, followed by LTFD4, LTFD1, and finally LTFD2.
Despite returning the highest Cllr, LTFD3 has the highest percentage of correct
SS and DS comparisons. It also offers the lowest EER, while providing the
strongest strength of evidence for SS and DS. The suspected reason for Cllr
being at its highest for LTFD3 is that Cllr appears to be greatly affected by
parameters (e.g. vowels) that produce wider ranges and higher magnitudes of
LLR. While producing these correct SS and DS comparisons with significant
strengths of evidence, it also tends to cause comparisons to yield incorrect SS
and DS comparisons with high strengths of evidence. For this reason, high Cllrs
appear to be being calculated for parameters that have the potential to offer
more in terms of correctness and the magnitude of the strength of evidence.
This was seen in Chapter 4, where although AR performed very poorly as a
discriminant, it achieved a lower Cllr than LTFD3, because the magnitude of the
166
AR LLRs overall were smaller. The same is true of LTFD2, which has the lowest
Cllr and also the smallest magnitude LLRs33.
Individually, LTFD1-4 performed relatively well, but the combination of
the formants can potentially yield even better performances. Table 5.4 below
follows the same structure as Table 5.3.
Table 5.4: Summary of LR-based discrimination for different LTFD1-4 combinations (100
speakers)
LTFD1+2 SS 70.0 0.417 -2.472 2.761
.2041 .7648
LTFD1+2 DS 85.0 -7.477 -76.391 1.996
LTFD2+3 SS 76.0 0.334 -7.828 3.768
.1392 .9630
LTFD2+3 DS 89.9 -14.173 -156.130 1.956
LTFD1+2+3 SS 74.0 0.625 -7.632 3.676
.1147 1.0161
LTFD1+2+3 DS 94.3 -19.307 -155.807 3.007
LTFD1+2+3+4 SS 84.0 1.160 -5.292 5.466
.0414 .5411
LTFD1+2+3+4 DS 97.4 -29.228 -162.931 2.854
Four different LTFD combination scenarios are presented in Table 5.4. LTFD1+2
performed the worst with respect to EER (.2041), but was the best in terms of
Cllr (.7648). LTFD1+2+3+4 performed the best with respect to EER, which was
.0414, and had the lowest Cllr (.5411). The highest proportion of correct SS and
DS comparisons was also returned by LTFD1+2+3+4, with 84% and 97.4%,
respectively. LTFD1+2 had the lowest proportion of correct SS and DS
comparisons with 70% and 85%, respectively. LTFD1+2+3 performed better
than LTFD2+3 with higher proportions of correct DS comparisons, mean LLR,
and EER. Overall, the combination of LTFD1+2+3+4 outperformed the other
three combinations as defined by EER and the proportions of correct SS and DS
comparisons.
33An additional explanation for the poor Cllr values could be that the system is not optimally
calibrated (see 8.5.3) as was also seen in 4.6.2.1.
167
The Tippett plots for the four LTFD1-4 combinations are presented in
Figures 5.17-5.20 below.
0.9
0.8
0.7
Cumulative Proportion
0.6
0.5
0.4
0.3
0.2
0.1 Different speaker comparisons
0
-80 -70 -60 -50 -40 -30 -20 -10 0 10
Log10 Likelihood Ratio
Figure 5.17: Tippett plot of LTFD1+2
168
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2 Same speaker comparisons

0.1
0
-160 -140 -120 -100 -80 -60 -40 -20 0 20
Figure 5.18: Tippett plot of LTFD2+3
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2 Same speaker comparisons

0.1 Different speaker comparisons
0
-160 -140 -120 -100 -80 -60 -40 -20 0 20
Figure 5.19: Tippett plot of LTFD1+2+3
169
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-180 -160 -140 -120 -100 -80 -60 -40 -20 0 20
Figure 5.20: Tippett plot of LTFD1+2+3+4
Figures 5.17-5.20 illustrate the range of LLR values calculated for the four
LTFD1-4 scenarios. LTFD1+2+3+4 had the largest positive values for SS LLR
and the largest negative values for DS LLR (i.e. best strengths of evidence in
both cases), and the best overall mean LLRs (SS = 1.16, DS = -29.228). LTFD1+2
had the smallest LLR ranges for both SS and DS, and the weakest mean DS LLR.
LTFD2+3 yielded the weakest mean SS LLR, and the second-weakest mean DS
LLR. LTFD1+2+3+4 offered the strongest LLR for same-speaker pairs at 5.466
(very strong evidence), and the strongest minimum LLR for DS at -162.931.
Overall, LTFD1+2+3+4 was the best combination of formants, followed by
LTFD1+2+3, LTFD2+3, and finally LTFD1+2. In comparison to the figures for
LTFDs for individual formants, the four combination scenarios were able to
significantly lower EER and improve the proportion of correct SS and DS
comparisons.
170
5.4.3 Results of Package Length
LTFDs were measured every 5msec in the current study; however, to
test the discriminant power of LTFD using MVKD34, multiples of 5msec
packages35 were created from multiple LTFD measurements to create localized
LTFD tokens (which are equivalent to short-duration portions of the recording).
The length of the package over which a distribution is calculated can vary. A
package length of 0.5 seconds was chosen for the study, as it was found to yield
the lowest EERs. The effects of package length variation can be seen in Table 5.5
for the LTFD combination of LTFD1+2+3+4. The size of the package length is
provided in the first column, followed by the percentage of SS and DS
comparisons that were correct. The mean, minimum, and maximum LLRs for SS
and DS comparisons are found in columns four through nine, while EER and Cllr
are provided in the last two columns.
Table 5.5: Package length variability
Mean Mean Min Max Max

Package SS % DS % Min DS
SS DS SS SS DS EER Cllr
Length Correct Correct LLR
LLR LLR LLR LLR LLR
.25 sec 76 97.96 0.90 -35.58 -6.51 -199.61 5.64 2.93 0.043 0.775
.5 sec 84 97.43 1.16 -29.23 -5.29 -162.93 5.47 2.85 0.041 0.541
1 sec 88 96.73 1.34 -24.17 -4.18 -134.33 5.29 2.79 0.042 0.400
2.5 sec 94 95.76 1.52 -17.67 -3.17 -98.41 4.99 2.88 0.043 0.281
5 sec 96 94.82 1.60 -13.85 -2.45 -84.60 4.77 2.90 0.042 0.239
10 sec 98 92.78 1.55 -9.27 -2.59 -62.94 4.41 2.68 0.056 0.257
The results in Table 5.5 suggest that package length affects the discriminant
performance of LTFD. An increase in package length corresponds to an
improvement in correct SS comparisons and Cllr, while there is a decrease in
34 LTFDs in their raw form readily lend themselves to a UBM-GMM algorithm for calculating
LRs. However, in order to test the MVKD formula, LTFDs were put into packages, as the MVKD
formula is not equipped to handle streams of data.
35 Package length was also used by Moos (2010) to determine stability within LTFD over
varying quantities of data. The packages are used in a similar way here, but are evaluated in
terms of their validity.
171
correct DS comparisons. In general, EER improves with the decrease in package
length; however, EER appears to have a threshold around 0.5 seconds, above
which it no longer improves.
It is important to note that these results apply only to the given study,
where the total duration of material per speaker was around 50 seconds. It
could be the case that package length has different effects on speech samples
longer or shorter than the 50 seconds used in the current study. However, this
analysis serves as a starting point for further investigation into the effects of
variability of package length and the overall length of speech samples on LTFD
results.
5.5 Discussion
The following section considers the discriminant value of higher formants and
lower formants, and also compares the results from the present study with
those from previous studies that used GMM-UBM-based LR calculations (Becker
et al., 2008; French et al. 2012).
5.5.1 Discriminant Value of Higher Formants
The results from individual LTFD LRs revealed that LTFD3 and LTFD4
performed better than the lower formants, LTFD1 and LTFD2, in discriminating
between speakers. These results suggest that the higher formants carry more
speaker-discriminatory information than the lower formants (also seen in
Jessen, 1997; McDougall, 2004; Moos, 2010; Simpson, 2008; Clermont et al.,
2008; Hughes, 2013). Table 5.6 provides an overview of previous studies
172
investigating the discriminant ability of formants where F3 (a higher formant)
was also found to outperform lower formants (F1 and F2).
Table 5.6: Overview of discriminant formant studies where F3 performs best
Formants Most
Study Data Measurements
Considered Discriminant
Jessen (1997) German; 20 speakers F1-3 Peaks in spectra F3
Australian English; 5
McDougall (2004) F1-3 Dynamic F3
speakers
Moos (2010) German; 71 speakers F1-3 LTFD F3
Simpson (2008) temporal
British English; 25
and Clermont et F1-3 midpoint of F3
speakers
al. (2008) formant
British English; 97
Hughes (2013) F1-3 Dynamic F3
speakers
The explanation for the better performance of higher formants than
lower formants (as seen in this study, and those listed in Table 5.6) can be
obtained by recourse to phonetic theory. The first and second formants are
responsible for encoding phonetic content (Ladefoged, 2006), where (lower)
frequencies are related in large part to tongue position: the first formant
correlates inversely with tongue height and the second formant is associated
with tongue frontness/backness (Clark and Yallop, 1990, p. 268). The range of
F1 and F2 values a speaker produces will be relatively constrained by the size
and shape of his/her vocal tract, while the given configuration of a speakers
vocal tract will determine its F1 and F2 values. In general, the lower formants
(i.e. F1 and F2) do not encode speaker-specific information; rather, they are
responsible for conveying phonetic content.
Contrastively, higher formants (specifically F3 and F4) have been
identified as encoding speaker-specific information, which makes sense given
that they are less affected by behavioral and physiological variation than are
lower formants (McDougall, 2004). This is because F3 and F4 are associated
173
with the resonances in the smaller cavities of the vocal tract, which allow for
less intra-speaker variation (i.e. smaller cavities offer a smaller space in which
resonances are produced; Peterson, 1959). However, the inter-speaker
variation of F3 and F4 is limited with respect to variation in the size of the vocal
tract (which does not show a wide range of variation; Xue and Hao, 2006). It is
important to note that Stevens and French (2012) have shown F3 to be
correlated in part to voice qualities that involve the backing of the tongue body,
an articulatory setting which was adopted by the majority of speakers in the
accent group they studied (SSBE speakers). The same was also found for
speakers of American English, where post-vocalic rhoticity results in the
lowering of F3 (Alwan et al., 1997). This means that although F3 is in part
responsible for differences in voice quality, which is to a large extent speaker-
specific, to some degree F3 can also encode accent information, specifically that
associated with tongue-body orientation (e.g. retracted tongue-body and a
pharyngealized voice quality; Laver, 1994).
To this extent, the suggestion that higher formants carry more speaker-
discriminant information than lower ones is borne out in the current research,
and provides an argument in support of the good performance of LTFD3 and
LTFD4 in the present study.
5.5.2 Comparison of LTFD, MFCC, MVKD, and GMM-UBM Results
The results presented in the current study were calculated using the MVKD
formula. However, GMM-UBM has also been used on the same data (French et
al. 2012), and LTFD on German data (Becker et al., 2008). The MFCC results
(French et al., 2012) and the LR results from French et al. (2012) and Becker et
174
al. (2008) for LTFD are compared (Table 5.7) to the results found in the present
study.
Table 5.7: Summary of LR-based discrimination for LTFD and MFCC in the current study and
competing studies
LTFD1+LTFD2+LTFD3 LTFD1+LTFD2+LTFD3+LTFD4 MFCC

SS DS EER SS DS EER SS DS EER
Current 74% 94.3% .1147 84% 97.4% .0414 - - -
study
French et - - - 94% 97.4% - 100% 95% -
al. (2012)
Becker et - - .053 - - - - - -
al. (2008)
The LTFD results from all three studies are generally similar36 regardless of
whether GMM-UBM or MVKD was used. However, given that French et al.
(2012) and the current study are based on the same recordings, it would be
expected that SS comparison results were more similar than they are (94% to
84%, respectively). This could suggest that for LTFD it is preferable to use
GMM-UBM over MVKD. However, it is important to note that 100 SS
comparisons were made by French et al. (2012), whereas the current study only
conducted 50 SS comparisons. Therefore, it is plausible that this 10% difference
could be due in part to the disparity in sample size (10% is equivalent to five SS
comparisons).
The tendency (albeit a small one) is for LTFD to miss SS pairs, and for
MFCC to mistake DS pairs for SS pairs. In view of this, it could be argued that in
the context of security, where investigators are working to put together a list of
potential suspects, MFCCs would be the preferred analysis. This is because
MFCCs are more likely to include additional suspects (despite their innocence)
rather than miss them entirely. Additionally, MFCCs tend to over-estimate
36Becker et al. (2008) also included results using bandwidths. However, those results are not
presented here.
175
similarity when comparing non-similar speaker pairings. In a judicial context,
the opposite could be argued: LTFD analysis should be preferred, insofar as it is
less likely to result in innocent suspects being misidentified as criminals. This is
because LTFD has the tendency to be under-sensitive when it comes to
identifying a guilty suspect as the criminal (by making judgments of non-
similarity when the samples are in fact similar by virtue of having been spoken
by the same talker).
5.6 Conclusion
Overall, the results presented in this chapter suggest that LTFD is a good
speaker discriminant, despite all LTFDs having variance ratios that imply intra-
speaker variation that is higher than inter-speaker variation. The combinations
of LTFD1-4 in 5.4.2 achieved higher levels of discrimination than single
LTFDs. The best combination, LTFD1+2+3+4, had an EER of only 0.0414, which
is extremely low compared to those found for AR in Chapter 4. Following the
results of the survey reported in Chapter 3, it appears that experts were correct
in identifying formants (in one form or another; e.g. LTFD, or for individual
phonemes) as one of the most useful speaker discriminants.
A known limitation of LTFD results from the study by Moos (2010),
where the values of LTFD means were higher in read speech than spontaneous
speech. It appears that speaking style can have a large impact on LTFD results.
It is important to consider in casework whether there is enough material
available to work with, and whether the material in the suspect and criminal
recordings is comparable, before carrying out an LTFD analysis.
176
The most attractive aspect of LTFD may not be in its successful results,
but in the fact that LTFD is not correlated with a number of other parameters
(Moos 2010). Correlation is often a challenge when vowels are analyzed
individually and later there is a desire to combine results from those multiple
vowels. This often results in a scenario where certain vowels are inevitably
correlated, and following nave Bayes (the combination of evidence through
the multiplication of individual LRs only when pieces of evidence are mutually
exclusive; Kononenko, 1990) they cannot be considered together as evidence. A
simple solution to this problem is to average across all vowel phonemes to
produce a LTFD. The only drawback to this lies in the high level of
generalization that is entailed when all vowels are averaged, meaning that
idiosyncrasies in individual phonemes may be overlooked. It appears that both
LTFD and MFCC analysis can provide insights into the vocal tract; however,
under an LR framework only one of these vocal tract parameters (LTFD or
MFCC) would be combined with other pieces of speech evidence into an overall
LR (owing to the strong correlations between LTFD and MFCC; French et al.,
2012). For this reason, unless a single phoneme can yield more promising LR
results for different populations, these results suggest that LTFD should be
considered over individual vowel analysis under the LR framework.
177
Chapter 6 Long-Term Fundamental
Frequency
6.1 Introduction
Long-term fundamental frequency (F0) is a commonly-used feature in forensic
speaker comparisons. Clark and Yallop (2001, p. 332) define fundamental
frequency as the number of times per second that the vocal folds complete a
cycle of vibration. Long-term fundamental frequency is the measure of
fundamental frequency over longer segments of speech, instead of smaller
intervals (e.g. a phoneme, a word). Clark and Yallop further explain that F0 is
controlled by the muscular forces determining vocal fold settings and tensions
in the larynx, and by the aerodynamic forces of the respiratory system which
drive the larynx and provide the source of energy for phonation itself (Clark
and Yallop, 2001, p. 333). Speakers are known to differ from one another in the
distribution of spectral energy (of F0) within their speech, due largely to
anatomical reasons and the way in which individuals manage their
phonatory/vocal tract settings (Clark and Yallop, 2001). For this reason, F0 is
commonly analyzed in forensic speaker comparisons, with the aim of
identifying those speaker-specific differences found in vocal fold vibrations and
phonatory and other vocal settings. The survey completed by expert forensic
phoneticians discussed in Chapter 3 reports that all experts considered F0 in
casework. Alongside voice quality, F0 was also claimed to be the most useful
speaker discriminant by experts. The most commonly-measured aspects of F0
were mean and standard deviation ( 3.9.1.2).
178
Despite the popularity of F0, the parameter is not immune to exogenous
factors such as emotion, disguise, alcohol, drugs, telephone transmission, and
recording codecs (Braun, 1995; Gold, 2009; Knzel, 2001; Papp, 2008). It is
already highly variable within speakers and the presence of these factors makes
it even more so. Regardless of this, however, the experts expectations remain
that it is useful in discriminating between speakers. An example of F0 playing a
key role in a forensic case is outlined in Nolan (1983, p. 124). The expectation
of F0 being a good speaker discriminant may stem from the view that it is to
some extent anatomically determined (Hudson et al. 2007 p. 1809). On a
positive note, F0 has been shown to be rather robust to background noise and is
not greatly affected by telephone transmission (Braun, 1995).
In order to evaluate experts expectations regarding the discriminant
power of F0 (expressed in terms of mean and standard deviation), empirical
testing is required on large homogeneous groups of speakers. There have been
a number of studies examining F0 in English (Hudson et al., 2007; Graddol,
1986; Loakes, 2006). However, only Hudson et al. (2007) provides statistics for
a group of English speakers. There is also only one study on English that reports
on within-speaker variability. However, this is for a set of only eight speakers of
Australian English (Loakes, 2006). For this reason, the current chapter
examines inter- and intra-speaker variation in F0, and considers the
discriminant potential of F0.
Fundamental frequency has previously received a large amount of attention in
forensic phonetic research. Kinoshita et al. (2009, p. 92) suggest that the
179
popularity of F0 stems from promising results in early speaker recognition
research (such as the work by Atal (1972)), as well as F0 fitting three of
Nolans standards for good forensic speaker comparison features its
robustness, measurability, and availability (Nolan, 1983). This section brings
together relevant F0 literature, and divides it into two parts. The first part
considers F0 in general as a population statistic and a speaker discriminant in
forensic phonetics, and the second examines exogenous factors that can affect
F0 and its measurement.
6.2.1 F0 as a Speaker Discriminant in Forensic Phonetics
Research on long-term F0 has resulted in a number of published statistics that
are often cited as reference data, especially in relation to forensic speaker
comparison casework and research. Fundamental frequency statistics for male
and female speakers in both read and spontaneous speech, and across multiple
languages, are provided in Traunmller and Eriksson (1995). Many new studies
have been conducted since Traunmller and Eriksson (1995). The majority of
those that have had forensic motivations have specifically analyzed F0 in
spontaneous speech, and among larger and/or more homogeneous groups of
speakers.
Rose (2003) reports long-term F0 measurements for non-
contemporaneous read speech (recordings separated by approximately one
year), produced by six male speakers of Australian English. The six speakers
had a mean F0 in the first recording of 113.6Hz (range: 101.9-124.8Hz) and a
mean standard deviation of 21.7Hz (range: 15.24-30.5Hz). The second
recording had a mean F0 of 114.5 Hz (range: 101.4-127.6Hz) and a mean
180
standard deviation of 17.4Hz (range: 14.3-19.2Hz). Loakes (2006) also presents
F0 values for male speakers of Australian English. Measurements of F0 were
taken at the midpoints of vowels in eight-minute recordings. The eight speakers
had a mean F0 of 105.2Hz, a median F0 of 103.1Hz, a mode F0 of 107Hz, and a
mean standard deviation of 16.4Hz. The F0 values for spontaneous speech in
Loakes (2006) were lower than those reported by Rose (2003). However, read
speech tends to elicit higher F0 values than spontaneous speech (Loakes, 2006).
Lindh (2006) reports long-term F0 values for 109 young male Swedish
speakers taken from short samples of spontaneous speech. The male speakers,
aged 20-30, had a mean F0 of 120.8Hz, a median F0 of 115.8Hz, and an average
alternative baseline F0 of 86.3Hz. The alternative baseline is the value (in Hz)
that falls 7.64% below the mean F0 (approximately 1.43 standard deviations;
see Lindh (2006) for more on alternative baseline). These F0 values are higher
than those found for Australian English. Rose (2002) suggests that F0 values
may be language-specific. The findings presented by Lindh (2006) also indicate
that collecting F0 statistics is necessary for different languages and perhaps
even different dialects/accents in order to understand the significance of
specific F0 measurements in forensic casework.
The study most relevant to the research reported in the current chapter
is that by Hudson et al. (2007), which investigates long-term F0 in the speech of
100 male speakers of British English drawn from the DyViS database. The
authors use Task 1 of DyViS, where individuals are taking part in a simulated
police interview. Three to five minutes of spontaneous speech per speaker were
analyzed after all background noises were removed. A Praat script was used to
extract mean, median, and mode F0 for each speaker. The aim of the research
181
was to gain an understanding of the distribution of F0 in a large homogeneous
group of English speakers primarily for forensic casework. Hudson et al. (2007)
report a group mean F0, median F0, and mode F0 of 102.2, 106, and 105 Hz
respectively. The study provides the population results in a format that is useful
for interpreting data with respect to between-speaker variation. However, it
does not offer any analysis of within-speaker variation. Under a numerical LR
framework it is necessary to provide an analysis of intra-speaker variation in
addition to the more commonly-studied inter-speaker variation. As such,
Hudson et al. (2007) cannot attach a numerical value to the level of
discrimination that can be achieved using only F0 as a discriminant.
The studies detailed above have reported average F0 values for groups
of speakers, while ignoring individual speaker variation (aside from Loakes,
2006 and Rose, 2003, but at a very limited level). Kinoshita (2005) was the first
to investigate intra- and inter-speaker variation in F0 in conjunction with the
discriminant power of F0 on a large scale. Kinoshita provides long-term F0
statistics derived from non-contemporaneous samples of spontaneous speech
for 90 male speakers of Japanese, reporting a mean F0 of 135.7Hz and a
standard deviation of 26.4Hz. Likelihood ratios were calculated using Lindleys
(1977) formula and synthetic (i.e. invented) criminal and suspect F0s (both
mean and SD) were created. The 90 male speakers acted as the reference
population. The results presented had a small range of LR estimates, and were
rather close to unity (i.e. not supporting a preference for one hypothesis or the
other). Kinoshita (2005) therefore suggested that long-term F0 is not a very
strong speaker discriminant and that it contributes very low strength of
evidence.
182
6.2.2 Effects of Exogenous Factors on F0
Intra-speaker variation is caused by numerous factors, such as
intoxication, emotion, vocal effort, disguise, and recording/transmission
technology, to varying extents (Braun, 1995; Gold, 2009; Junqua, 1996; Knzel,
2001; Linard and Di Benedetto, 1999; Papp, 2009; Zetterholm, 2006). F0 has
been shown not to be as robust as previously believed, and it is sometimes the
case in forensic cases that F0 measurements are not robust to the effects of
these exogenous factors (Peter French, p.c.).
Braun (1995) draws attention to a number of factors known to affect F0
and which may have some relevance to forensic situations. An exhaustive list of
known effects on F0 is presented with relation to technical, physiological, and
psychological factors. She gives, as examples of technical factors, tape speed,
electronic voice changers/disguise (Hollien and Michel, 1968; Knzel, 1987),
and sample size (French, 1990; Horii, 1975; Mead, 1974; Steffan-Battog et al.,
1970). Physiological factors affecting F0 include speaker race (Hudson and
Hollbrook, 1981), age, a history of smoking (Braun, 1994; Gilbert and Weismer,
1974; Murphy and Doyle, 1987; Sorensen and Horii, 1982), alcohol
consumption (Klingholz et, al. 1988; Knzel et al., 1992; Pisoni and Martin,
1989; Sobell et al., 1982), testosterone drugs and anabolic steroids (Bauer,
1963; Berendes, 1962; Damast, 1964; 1967), removal of cysts/nodules/polyps
(Bouchayer and Cornut, 1992), surgical stripping of the vocal folds after
edema/tonsillectomy/thyroidectomy (Ardnt, 1963; Fritzell et al., 1982;
Keilmann and Hlse, 1992), shortening of vocal folds (Oats and Dacakis, 1983),
and lingual block (e.g. use of anesthesia; Hardcastle, 1975). Finally, Braun
(1995) identifies a number of psychological factors known to affect F0, which
183
are emotions (sorrow, anger, fear; Williams and Stevens, 1972), stress (Hecker
et al., 1968, Scherer, 1977), vocal fatigue (Novak et al., 1991), depression
(Darby and Hollien, 1977; Hollien 1980; Scherer et al., 1976), schizophrenia
(Hollien, 1980; Saxman and Burk, 1968), time of day (Garrett and Healey,
1987), and background noise level (Dieroff and Siegert, 1966; Lombard, 1911;
Schultz-Coulon, 1975; Schultz-Coulon and Fues, 1976). These extensive lists
detailed by Braun (1995) pose problems that FSC experts are confronted with
when analyzing F0. Many other studies have been carried out since Braun
(1995) to examine the effects of external factors on F0. Those studies that are
most relevant to casework are detailed below.
Linard and Di Benedetto (1999) examined the effects of vocal effort on
F0. They looked at 12 French vowels spoken in isolation by ten speakers (five
males and five females). Vowels were repeated multiple times to an
experimenter who stood at varying distances in the room from the speaker
(close, normal, and far). The distance at which the experimenter stood relative
to the speaker was intended to induce change in the vocal effort that the
speaker assumed would be required for the experimenter to hear the speaker
clearly. Linard and Di Benedetto (1999) found F0 to increase by around 5Hz.
Voice disguise is another common cause of variation in F0. Knzel
(2000) reports that nearly 25% of cases in Germany involve voice disguise, and
he specifically investigated the effect of such disguise on F0. He analyzed read
speech from 100 speakers (50 males and 50 females) where they were asked to
adopt different voice disguises (high, low, and denasalized). Knzel showed that
speakers were effectively and consistently able to disguise their voices using F0
modulation, some to extreme levels. Most importantly, he notes that individuals
184
were observed applying different phonatory strategies, which caused difficulty
in associating the change in F0 with a particular change the speaker has made in
his/her phonatory setting. Zetterholm (2006) has also examined disguise and
imitation of voices in relation to changes in F0 using the speech of a single
individual (a professional impersonator), impersonating 12 different popular
Swedish TV personalities. She found that the impersonator was able to vary his
modal F0 voice from its normal average of 118Hz, to anywhere between 97Hz
and 225Hz.
The effects of recording and transmission technology on speech has
received more attention of late, perhaps due in part to the advent of new and
emerging technologies. This opens new avenues for potential technical effects
on F0 and on speech in general. Gold (2009) considers one such area, by
investigating the effects of video and voice recorders in cellular phones. Three
different cellular phones were used in the experiment. All phones encoded the
speech signal using an AMR or mpeg4 codec for voice and video recorders
respectively. A change in the speakers F0 was found to result in differences of
between one and five percent in mean F0, and the SD of F0 changing from 9 to
63.6% for a single cellular phone. Overall, voice recorders (AMR codec) were
found to make bigger changes in F0 than video recorders did. It has previously
been shown that the GSM AMR codec (a speech-encoding codec similar to the
AMR) has the tendency to change voiced frames into unvoiced frames, and vice
versa, which in turn affects F0 measurements (Guillemin and Watson, 2008, p.
216). This could be the case for the AMR codec found in cellular phones, and any
recording devices in general that incorporate a similar-functioning codec to the
GSM AMR.
185
Variation in F0 can be caused by numerous exogenous factors, many of
which (those relevant to casework, at any rate) have been detailed above. The
variability of F0 due to external factors introduces many difficulties when
comparing recordings that have been affected by different combinations of
exogenous factors. For this reason it is important to consider the effects of
within- and between-speaker variation when investigating the discriminant
potential of F0.
6.3 Population Statistics for Fundamental Frequency
The following section presents population statistics for F0 in a large,
linguistically homogeneous group of 100 male speakers. These data serve as the
first of their kind in providing detailed information on intra-speaker variation
in F0 among individuals who speak Southern Standard British English (SSBE).
Additionally, population data are provided for between-speaker variability in
F0, where in the discussion it will be made apparent that the variability is
similar to that found in Hudson et al. (2007).
6.3.1 Methodology
The current study uses the recordings from Task 2 of the DyViS
database. Each recording for all 100 speakers was used in its entirety. However,
after the editing of the files, the recordings were between 2:25 minutes and
11:17 minutes in length, with an average of 6:21 minutes per file. Using Praat
(version 5.1.35), multiple passes were made through the recordings to ensure
that there was only speech remaining. The first phase consisted of the removal
of all the portions where the interlocutor was speaking, and any silent pauses
186
with the talkers speech. The second phase removed all intrusive noises in the
recordings, including background noises, laughter, coughs, and sneezes. A final
listening phase was used to ensure that everything but the speech of the
speaker had been properly removed from the recordings. This final phase saw
only minor edits that typically amounted to the removal of less than one second
of net speech per speaker.
As the calculation of LRs requires multiple tokens per speaker, it was
necessary for all recordings to be divided into segments in order to establish
within-speaker variability. The amount of net speech necessary for a given F0
token has not been previously tested. For this study I chose to segment the
speech into 10-second intervals (see 6.4.3 for effects of package length). Each
file was then annotated using a Praat text grid. The tier represented the package
length (in seconds) according to which the speech signal was subsequently
segmented. Starting from the beginning of each recording, intervals were
marked out in the text grids until the end of each recording. If the final segment
did not meet the interval length requirement, it was not included in analysis.
Figure 6.1 depicts an example of the text grid annotations used for all
recordings.
187
Figure 6.1: Example of a text grid annotation
A Praat script (pitch_mean_results.txt; created by Henning Reetz, 2009) was
used to extract mean F0 and standard deviation values for each interval. The
Praat script was set to a frequency range of 50 300 Hz (following Hudson et al.
2007). After reviewing the F0 Praat picture distributions (for octave jumps and
unwanted pitch artefacts), 64 speakers were found to have reliable F0. The
remaining 36 speakers F0 values contained obvious errors (e.g. octave jumps)
and were therefore re-run using tailored ranges. The tailored ranges were
chosen through trial and error, where the range with the least amount of errors
was chosen as the best possible frequency range. These are detailed in Table
6.1.
188
Table 6.1: Tailored Frequency Ranges for Selected Speakers
Frequency Range (Hz) Speaker

50-150 024, 026, 050, 062
50-160 074, 081, 099
50-200 052, 056
55-160 007, 009, 025, 029, 036, 040, 073, 096
55-200 045, 049, 055, 065, 075, 078, 082
75-200 003, 015, 021, 035, 041, 058, 059, 066, 083, 092
75-160 006, 014
After the Praat script was re-run using tailored frequency ranges for these
speakers, all F0 means and standard deviations were imported into Microsoft
Excel for further analysis. There were a total of 7,447 intervals for all 100
speakers. On average each speaker had 74 intervals.
6.3.2 Results
The distributions of F0 mean and standard deviation for individuals are
presented in Figures 6.2 and 6.3. The y-axis represents the number of speakers
that fall within a given range and the x-axis depicts F0 in Hertz (Hz).
189
Mean Fundamental Frequency
18
14
12
10
8
6
4
2
0
Fundamental Frequency (Hz)
Figure 6.2: Distribution of mean fundamental frequency
There is a normal distribution in the DyViS corpus for mean F037, as illustrated
in Figure 6.2. The mean F0 for the population is 103.6Hz, with an overall mean
F0 range of 79.9Hz-136Hz. The standard deviation of the means is 12.77Hz.
There are no suspected outliers (as defined in Chapter 4) in the mean F0 data.
Technically, this is the mean of the means of the means (i.e. the mean across speakers of the
37
means across tokens of each speaker of the means of all the raw F0 values of each token).
190
Mean Standard Deviation for
Fundamental Frequency
30
25
Number of Speakers
20
15
10
0
6-8 10-12 14-16 18-20 22-24 26-28 36-38
Fundamental Frequency (Hz)
Figure 6.3: Distribution of standard deviation in fundamental frequency
There is a roughly normal distribution for the SDs of F038 in Figure 6.3, with a
slight positive skewing due to a number of outliers. There are seven suspected
higher outliers (22.62Hz (for two speakers), 23.25Hz, 23.29Hz, 24.82Hz,
25.59Hz, and 27.62Hz) and one extreme outlier at 37.32Hz. Including the
outliers, the mean SD for the population is 15.1Hz, with a range of 7.4Hz-
37.3Hz. If those outliers are removed, the distribution becomes more normal
and the mean SD is then 14.17Hz.
The cumulative distribution graphs of mean F0s and F0 SDs in Figures
6.4 and 6.5 (respectively) show the percentiles at which a given F0 mean or SD
falls in relation to the population. The y-axis represents the cumulative
38Technically, this is the mean SD of the mean SD (i.e. the mean SD across speakers of the SDs
across tokens of each speaker of the means of all the F0 raw values of each token).
191
proportion of the population, and the x-axis presents fundamental frequency
(Hz).
Mean Fundamental Frequency

100
90
80
70
60
50
40
30
20
10
0
75 80 85 90 95 100 105 110 115 120 125 130 135 140
Frequency (Hz)
Figure 6.4: Cumulative percentages for mean fundamental frequency
The curve in Figure 6.4 is characterized by a steep central section but has gentle
gradients at both ends. The data show that the lowest 20% of the speakers have
a mean F0 below 93Hz, while the highest 20% have an F0 above 115Hz. This
leaves only a narrow band of 22Hz in which the remaining 60% of speakers are
found. This is indicated by the steepest portion of the trajectory.
192
Mean Standard Deviations for
Fundamental Frequency
100
90
80
70
60
50
40
30
20
10
0
7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39
Figure 6.5: Cumulative percentages for standard deviation in fundamental frequency
The first half of the curve in Figure 6.5 is characterized by a steep trajectory,
while the second half has a fairly long gradient trajectory at the upper end. The
lengthy gradient to the end of the trajectory is due to the suspected outliers and
extreme outliers confirmed above in the current section. Observing the spread
of the SDs, the lowest 20% of speakers have F0 SDs below 12Hz, and the highest
20% have F0 SDs above 17.5Hz. This leaves a remarkably narrow band of 5.5Hz
in which the majority of speakers fall (60%). Overall, the F0 data have a
variance ratio of 0.7152, which indicates that there is more variation occurring
within speakers than between them.
6.3.3 Discussion
The results for the present study are very similar to those reported by
Hudson et al. (2007), which used Task 1 of DyViS. The difference in the two
193
group means is only 2.4Hz. Results from the present study are also similar to
those presented in Loakes (2006), who gave a mean F0 of 105.2Hz for the
spontaneous speech of Australian English-speaking males. For non-English
languages the results presented in 6.3 are somewhat dissimilar in that both
Swedish (Lindh, 2006) and German (Knzel, 1989) report higher mean F0s at
115.8Hz and 120.8Hz, respectively.
The lower mean F0 values found for SSBE compared to those found for
other languages could in fact be a result of F0 being language-specific, as
suggested by Rose (2002). However, it could potentially be caused by the fact
that numerous speakers in the DyViS database have creaky voice qualities
(Hudson et al., 2007), and the creaky voice qualities of speakers were included
in the current study. This is because the F0 of speakers who use creaky
phonations a lot tend to result in bimodal distributions, with the first peak
representing the creaky voice quality and the second peak representing modal
phonation. In order to calculate a mean F0 for a speaker the two phonation
types are averaged, which thus results in a lower mean F0 (Hudson et al., 2007).
As such, it is most likely the case that, as pointed out in Hudson et al. (2007),
speakers modes did not correspond to their means. For this reason it would be
ideal to find a more accurate way of representing the mean F0 of creaky-voiced
individuals. Overall, the results also suggest that mean F0 and SD are perhaps
not the best measures for those speakers with intermittently-present creaky
voice.
194
The discriminant results for F0 are presented in the following section. The first
part provides LR results for F0, and the second part considers the effects of
package length on LR results for F0.
6.4.1 Methodology
As seen in 4.6.1, likelihood ratios were calculated using a MatLab
implementation of Aitken and Lucys (2004) Multivariate Kernel-Density
(MVKD) formula (Morrison, 2007). An intrinsic methodology was used,
whereby the test and the reference speakers came from the same population of
100 speakers. Speakers 1-50 were used as the test speakers, while speakers 51-
100 served as the reference speakers. Mean F0 and SD parameters were both
used for each token spoken by a given individual in the calculation of the LRs.
Performance of the system was assessed in terms of both the magnitude of LRs
(Champod and Evett, 2000) and system validity (Cllr and EER).
6.4.2 Results for F0
The results for the calculation of LRs for F0 are summarized in Table 6.2.
The second row contains the results from same-speaker (SS) comparisons and
the third row contains the different-speaker (DS) comparison results. The
second column indicates the percentage of comparisons in which speakers were
correctly identified, whereby a log likelihood ratio (LLR) above zero was correct
for a SS comparison and an LLR of less than zero was a correct judgment for DS
comparisons. The mean LLR is found in the third column, followed by the
minimum and maximum LLRs. The final two columns present the performance
of the system in terms of EER and Cllr, respectively.

195
Table 6.2: Summary of LR-based discrimination for F0 (100 speakers)
Comparison % Correct Mean LLR Min LLR Max LLR EER Cllr
10 sec SS 92.0 0.958 -3.404 1.936
0.0849 0.4547
10 sec DS 89.9 -24.204 -269.159 1.906
Table 6.2 shows that SS comparisons slightly outperform DS comparisons in the
percentage of correct judgments. The mean LLR for DS offers very strong
evidence to support the defense hypothesis (Hd; Champod and Evett, 2000),
while the mean LLR for SS only offers limited evidence to support the
prosecution hypothesis (Hp). Even the Max LLR for SS does not reach a strength
of evidence of 2 (instead, only moderate evidence to support Hp is indicated by
the value of 1.936). The EER for the system is higher than that found for a
combined LTFD system in Chapter 5 (0.0414; see Table 5.4), but is significantly
better than that found for AR in Chapter 4 (0.334; see Table 4.7). The Cllr for F0
as a system is generally better than the Cllrs achieved in Chapters 4 and 5 for
AR and LTFD. A Cllr closer to zero would nonetheless be desirable.
The Tippett plot in Figure 6.6 offers a visual measure of the performance
of F0 as a discriminant feature.
196
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-180 -160 -140 -120 -100 -80 -60 -40 -20 0 10
Figure 6.6: Tippett plot of fundamental frequency
Figure 6.6 shows that there is a narrow range in LLR for SS and that most LLRs
for SS are relatively similar. The DS comparisons have a wider spread of LLRs. It
is also clear that DS comparisons can achieve very large LLRs, which offers a
high strength of evidence.
6.4.3 Results of Package Length
In order to establish variability within a speaker under an LR
framework, multiple tokens of a speech parameter of an individual are needed
for analysis. This involves dividing the recording into multiple sections (or
tokens). The most efficacious token length (or referred to here as package
length) has not been previously established. Therefore, a package length of 10
seconds was chosen for the study as it was found to yield the lowest EER.
However, it is possible to vary the size of the package length (similar to that
seen in Chapter 5). The effects of package length variation can be seen in Table
197
6.3 for F0. The table has the same formatting as that of Table 6.2, although Table
6.3 is expanded to include additional rows containing the various package
lengths (5, 10, 15, and 20 seconds).
Table 6.3: F0 package length variability for LR results
5 sec SS 88.0 0.868 -5.176 2.030
0.1010 0.5634
5 sec DS 91.1 -29.879 -308.588 2.031
10 sec SS 92.0 0.958 -3.404 1.936
0.0849 0.4547
10 sec DS 89.9 -24.204 -269.159 1.906
15 sec SS 92.0 0.970 -2.526 1.999
0.1016 0.4407
15 sec DS 89.0 -20.964 -233.785 1.809
20 sec SS 92.0 0.960 -2.536 1.880
0.0967 0.4383
20 sec DS 88.7 -18.620 -206.961 1.717
The results in Table 6.3 suggest that package length affects the discriminant
performance of F0. However, the increase of package length does not appear to
be linearly correlated with the overall system performance in terms of EER. For
Cllr, there does appear to be a direct relationship between the increase in
package length and the improvement in Cllr.
Effects of package length can also be considered from inspection of the
values shown in Table 6.4, which displays the results for mean F0s, F0 range,
standard deviations (SD), and range of SDs across the four different package
lengths.
Table 6.4: Fundamental frequency across different package lengths
Package Mean of Means Range of Means Mean of SDs Range of SDs

Length (Hz) (Hz) (Hz) (Hz)
5 sec 103.2 79.7-136.1 14.3 6.9-36.2
10 sec 103.6 79.9-136.0 15.1 7.4-37.3
15 sec 103.3 79.8-136.2 15.4 7.6-37.6
20 sec 103.2 79.9-136.1 15.5 7.8-37.9
198
Table 6.4 shows that there is relatively little difference between F0 results as a
function of the different package lengths. The biggest difference found in the
results is in the mean of SDs for 5 seconds (14.3Hz), compared to the mean of
SDs for the larger package lengths (15.1-15.5Hz). The 5-second package length
was also found to have the biggest difference in Cllr (Table 6.3) with the longer
package lengths. On the basis of these results, it could be argued that choosing a
package length of 10 seconds or above will give an accurate representation of
the data.
It is important to note that like the package length results found for
LTFD in Chapter 5, the results presented in this section relate specifically to the
present recordings, in which the total length of material per speaker was
around six minutes. It could again be the case that package length affects longer
or shorter speech samples differently. However, this analysis serves as a
starting point for further investigation into this issue.
6.5 Discussion
The results presented in the present study provide a starting point for further
investigation into the discriminant value of F0. However, the study was limited
by the highly controlled nature of the recordings, which were relatively free
from the influences of the exogenous factors that are known to affect F0 values,
as detailed in 6.2.2. More studies which incorporate those factors are needed.
The results of the present study were produced using only mean and
standard deviation as discriminant parameters of F0. This choice was dictated
by opinion given in the survey completed by expert forensic phoneticians. The
survey also reported that it is not uncommon for experts to use other measures
199
of F0 in their casework. Kinoshita et al. (2009) showed promising results when
using parameters that described the distribution of F0 more precisely (skew,
kurtosis, modal F0, and modal density). This points to a need to reassess the
measures commonly used in relation to F0. Different measures of F0 could lead
to more detailed descriptions of F0 distributions that also achieve a higher
6.6 Conclusion
The results presented in this chapter suggest that F0 is a moderately good
speaker discriminant overall, and has promise for demonstrating that two
voices have come from the same speaker (rather than different speakers) in the
same recording (achieving an EER of 0.0849). However, it is not known how
well F0 can discriminate between individuals when same-speaker evidence
comes from different recordings. Previous literature would suggest that its
discriminant potential will decrease when same-speaker evidence from
different recordings is introduced (see 6.2.2). F0 as a speaker discriminant
showed more variation occurring within speakers than between speakers.
However, F0 does show there to be more variation present between speakers
than do AR (Chapter 4) and individual LTFDs (Chapter 5).
It is difficult to ascertain whether experts responding to the survey were
correct in identifying F0 as a good speaker discriminant. Results suggest that
they are correct in that F0 does well discriminating same speakers that come
from the same recording, but it is uncertain whether that result will hold true
when same-speaker comparisons involve different recording sessions or
introduce degrading factors (e.g. disguise, intoxication, background noise). The
200
good recording conditions and high audio quality used for the current study are
not reflective of those found in real casework.
Owing to the many exogenous factors detailed in 6.2.2, the mere
comparison of mean F0s and SDs is unlikely to advance the methods used for
the speaker comparison task dramatically on its own. However, as always,
exceptions are to be made for those individuals who lie towards the margins of
the distribution curve or who can be classed as outliers, and the case remains
for using F0 in conjunction with other speech parameters.
201
Chapter 7 Click Rate
7.1 Introduction
The survey results presented in Chapter 3 indicated that experts examine many
non-linguistic features as part of their analysis in FSCs. Those non-linguistic
features can include patterns of audible breathing, laughter, throat-clearing,
tongue clicking, and both filled and silent hesitation phenomena. In respect of
tongue clicking, Chapter 3 shows that 57% of the practitioners questioned
examined recordings for the presence of velaric ingressive stops (i.e. clicks),
and 18% considered them to be a highly discriminant feature.
Research into the discriminant ability of parameters in forensic speech
science has focused primarily on vowels, and to some extent consonants and
fundamental frequency (Gold and Hughes, 2013). However, there remains a gap
in the literature pertaining to the discriminant ability of non-linguistic
parameters (e.g. clicks).
This chapter investigates the speaker discriminant power of clicks,
which are defined here as a linguistic parameter rather than a non-linguistic
parameter (reported in Chapter 3 as non-linguistic). This is because the clicks
analyzed in this chapter are used by speakers in a discursive manner that can be
classified as conveying linguistic meaning (i.e. they are used here as a discourse
marker in conversation). The first part of this chapter investigates the
discriminant power of clicks by analyzing population statistics for click rate,
and the second portion analyzes the robustness of clicks in relation to
accommodation effects. The final limitation section in this chapter is devoted
entirely to discussion of calculating likelihood ratios for clicks, and the
202
difficulties in attributing a numerical strength of evidence to measure discrete
data.
Ladefoged (2006, p. 292) defines a click as a stop made with an ingressive
velaric airstream, such as Zulu [ . Laver (1994, p. 174) explains further that a
major ingredient in the production of the airstream [for clicks] is a complete
closure made by the back of the tongue against the velum. A second closure is
also made, further forward in the mouth, either by the tip, blade or front of the
tongue or by the lips. For a lingual click, there is a closure made by the back of
the tongue coming into contact with the soft palate, and the front portion of the
tongue is then drawn downwards. This process increases the volume of the
space occupied by the air trapped in between the two closures rarefying the
intra-oral air-pressure. When the more forward of the two closures is released,
the outside air at atmospheric pressure flows in to fill the partial vacuum
(Laver 1994, p. 174). It is at this point that a click is realized. Figure 7.1 below
illustrates the actions of the vocal organs involved in the production of a click
sound.
203
Figure 7.1:) The action of the vocal organs in producing a velaric ingressive voiceless dental
click k (a) first stage velic and anterior closure; (b) second stage expansion of the enclosed
oral space; (c) third stage release of the anterior closure. (Laver, 1994, p. 176)
Figure 7.1 provides an illustration of the process involved for the vocal organs
in the production of a dental click. This is just one of six possible places of
articulation for clicks as recognized by the International Phonetic Association
(IPA). The five different click types are provided in the figure below, which is an
extract from the IPA chart.
Figure 7.2: IPA Chart - Clicks Excerpt
204
The six different clicks presented in Figure 7.2 above are most commonly
recognized for their existence in a number of African languages (Ladefoged
2006, p. 139) and extensive research has been carried out to document clicks in
those languages (e.g. Greenberg, 1950; Herbert, 1990; Jessen and Roux, 2002).
In African click languages, such as in Xhosa, Zulu, Sandawe, Hadzapi, Bushman,
Nama, !X, and !X clicks are used phonemically (Laver, 1994, p. 174). Clicks
are also found in English, but unlike those in African languages they are not
used phonemically. According to literature on clicks found in English, they have
typically been described as functioning on only a paralinguistic level to denote
the attitudes, intentions (e.g. encouraging a horse to move), and emotional
states of a speaker. Previous research suggests that certain clicks are used to
convey such things as annoyance (Abercrombie, 1967, p. 31; Ball, 1989, p. 10),
sympathy (Gimson, 1970, p. 34), and disapproval (Crystal, 1987, p. 126). There
is also evidence to suggest that the phonetic properties of clicks can vary
depending on their functions in English (Gimson, 1970, p. 34).
Wright (2005; 2007; 2011a; 2011b) presents an extensive amount of
research focused on clicks from a non-paralinguistic point of view, specifically
from a conversation analysts view. Wright proposes three different
classifications of click used in English conversation to index different meanings.
The first type are clicks that occur in the onset of a new sequence, the second
are clicks used in the onset of a new and disjunctive sequence, and the third
type are clicks produced in the middle of a sequence of talk when the speaker
is engaged in the activity of searching for a word (2005 p. 176). The following
are three examples from Wright (2005) of click types:
205
Fragment 1: Holt.SO.88.1.2/bath/
01: Bil: Hello:
02: Gor: Hi Bill
03: Bil: Hi Gordy
04: Gor: [] uh:m (0.4) are you going tonight
05: (.)
06: Bil: mm
07: Gor: .hhh (0.2) would you mind giving me a lif[t
08: no thats alright
Fragment 2: Holt.1.8/Saturday/
01: Les: so he had a good inni:ngs did[n't he
02: Mum: [I should say so: yes
03: (0.2)
04: Mum: marvellous
05: Les: [!]. .hhh anyway we had a very good evening o:n saturday
06: (0.2)
07: Mum: Ye:s
Fragment 3: Holt.U.88.2.2/natter/
01 Les .hhhh and theres the- the natte- uhm (0.2) (0.3) !
02: oh what's it called the natterjack's not so good now
Fragment 1 is an example of a click being used to start a new sequence, as noted
by the bilabial click in line 4. A second type of click is used for the onset of a
new disjunctive sequence in Fragment 2, which is illustrated by the alveolar
click on line 5. And the final click type is found in Fragment 3, where both
206
bilabial and alveolar clicks are being used to signify the search for a word by the
speaker in line 1. The three click types above, in combination with clicks being
used paralinguistically (i.e. to show emotion or affect), are used in the analysis
of click productions for this chapter.
In investigating clicks in respect of their speaker-discriminating
potential one can begin with the assumption that for any aspect of affect or
interaction management there is no homological function-form relationship.
For example, while one can signal annoyance, disapproval or sympathy by use
of clicks, there are many other ways of signaling these states to interlocutors.
Similarly, although clicks may be used to signal disjunction of conversational
topic or the fact that one is having difficulty finding a word, other forms
semantically empty sounds or lexical expressions can also fulfil these
functions. In other words, there is an element of speaker choice in the selection
of clicks over other possibilities in conveying emotive and attitudinal meaning
as well as in respect of topic organization and conversational turn management.
Given that this is so, one might reasonably expect there to be variability across
speakers in terms of whether clicks or other forms are their preferred option.
The possibility of such individual preferences provides a plausible theoretical
motivation for the observation made by the forensic practitioners surveyed in
Chapter 3 to the effect that clicks have high value as speaker discriminants.
However, while the proposition is credible and is no doubt based on
practitioners casework experience, it has not to date been subjected to formal,
empirical testing. The present chapter is an attempt to establish the speaker
discriminant value of one aspect of clicking behavior, namely frequency of
clicking, by such testing.
207
7.3 Data
The recordings analyzed were of 100 male speakers of SSBE aged 18-25 years
from the Dynamic Variability in Speech (DyViS) corpus (Nolan et al., 2009). Two
data sets were used, each of unscripted speech from a simulation of a
forensically-relevant situation. One (Task 1) was a mock police interview. Each
of the 100 speakers played the role of a criminal suspect and was interrogated
by one of two project interviewers (Int2 and Int3) who played the role of a
police officer investigating the interviewees supposed involvement in a crime.
The second set of recordings (Task 2) was of the subjects telephoning an
accomplice and explaining what had occurred in the police interview. The role
of the accomplice in this data set was played by the same project interviewer
(Int1) throughout. Although these were telephone conversations, the
recordings used for analysis were made at the subjects end of the line i.e. they
were of studio rather than telephone quality.
7.4 Methodology
For a feature to function as a good speaker discriminant, it must meet two
criteria: (a) it must vary (ideally quite widely) across speakers; (b) it must be
relatively stable within the speech production practices of individual talkers. In
this section, the methods employed to test the intra- and inter-speaker variation
of click rates are outlined. Task 2 recordings are used for the first portion of the
click analysis of the current study. As mentioned above, each speaker conversed
with a single interlocutor, Int1.
The first two minutes of each recording were ignored so as to allow for
speakers to settle into the interaction. All subsequent speech from each subject,
208
up to a maximum of five minutes net i.e. after excluding long pauses and the
speech of the interlocutor was then extracted and divided into one-minute
intervals (giving a combined total of 499 minutes of net speech). 99 of the 100
speakers produced enough speech to meet the five-minute target. One speaker
(speaker 012) fell just short of this, and the analysis was therefore based on
analysis of just four minutes of his speech. The extracted speech was examined
auditorily during two listening sessions in Sony Sound Forge (version 10.0;
analysis done auditorily) and Praat (version 5.1.35; auditory and acoustic
analysis done simultaneously) for instances of clicks. Any sounds that auditorily
and visually resembled clicks but were not apparently produced on a velaric
ingressive airstream were excluded from the analysis. This resulted in the
exclusion of 293 candidate sounds that were judged to be purely percussive39.
At the end of this process there were a total of 454 clicks left. Each click was
assigned to a functional category: either it functioned to convey affective
meaning, or fulfilled one of the interactional functions identified by Wright
(2007; 2011a; 2011b), i.e. initiating a new speaking turn, indicating topical
disjunction, or signaling that the speaker was searching for a word.
Illustrations of these interactional functions are provided in the following
transcribed excerpts from the recordings (clicks are indicated by the symbol !,
regardless of actual place of articulation):
39 Pike (1943 p. 103) says percussors differ from initiators in several ways in opening and
closing they move perpendicularly to the entrance of the air chamber . . . ; they produce no
directional air current but merely a disturbance that starts sound waves which are modied by
certain cavity resonators; they manifest their releasing or approaching percussive timbre only
at the moments of the opening and closing of some passage . . . Typical percussives are made by
the opening and closing of the lips, the tongue making closure at the alveolar ridge, the velum
closing, the vocal folds making a glottal closure, and the sublaminal percussive of the cluck
click (Ogden 2013 p. 302). The most common percussives found in the current data set were
related to the opening and closing of the lips.
209
Initiating new speaking turn
Int1: Hes a tour guide now you see
S007: ! Yeah yeah thats right and
Int1: Bear Pub
S011: Mhm
Int1: Do I know it
S011: ! Um it- near Harper Passage
Indicating topical disjunction
Int1: Um wha- did they trace that phone call when you were in the uh
grotty booth
S033: They asked me about it so I guess they probably have ! um but um
as I wasnt as I was telling them I didnt go through Parkville
Signaling word search
Int1: And um did you give her address
S086: Uh yeah I did
Int1: Just, you know, just refresh my memory
S086: Yeah, sorry on Dexter Road ! um in Dixon
Int1: Dixon this little village of Dixon
7.5 Results
Before addressing the central questions of inter- and intra-speaker variation,
some general findings on phonetic and functional aspects of the clicks are
presented.
210
7.5.1 Phonetic Properties of the Clicks
The scope of the present study did not extend to a detailed analysis of
the phonetic and acoustic properties of the clicks. However, in terms of their
place of articulation, approximately 95% were judged to be apical. With regard
to the passive articulator, they ranged from dental through alveolar to post-
alveolar. Through my own observation, dental clicks are characterized by a
longer and less well-defined release phase and by a higher-frequency center of
gravity and lower level of intensity than the other variants. At the other
extreme, post-alveolar clicks are the highest in intensity, have a relatively short
release and a greater concentration of energy at the lower frequencies. Without
wishing to prejudge the outcome of further work being undertaken on these
data, it appears that, at this stage at least, place of articulation proved very
difficult to classify more finely, and that no individual speaker clearly stood out
from the others in respect of this dimension. It is supposed that because clicks
are not used phonemically by SSBE speakers, the precise place of articulation
for clicks does not matter to a speaker or listener when used in sequence
management. It is rather that the presence of any form of apical click can signify
sequence management in conversation. Place of articulation, however, does
play an important role for those clicks used as affective markers, since place of
articulation for clicks has been shown to signify different emotions (Ball, 1989;
Crystal, 1987; Gimson, 1970).
7.5.2 Functional Aspects of the Clicks
The distributions of clicks against affective function and the three
interactional functions are represented in Figure 7.3.
211
Distribution of Click Types
250
200
Number of Total Clicks
150
100
50
0
word search turn initiation signaling topic affective
disjunction
Category of Click
Figure 7.3: Distribution of click occurrences by functional category
Of the 454 clicks that occur in the combined 499 minutes of speech examined,
word search accounts for just over half of all clicks (51.32%). Taken together,
turn initiation and disjunction signaling clicks represent a similar proportion
(48.24%) to those used to indicate word search. Affective use represents the
smallest category, with only two examples (0.44%). Whilst the latter may to
some extent be accounted for by the fact that the attitudinal stances that clicks
are used to convey (pity, disapproval) seldom arise in the type of conversation
represented in the DyViS recordings, it is nevertheless of interest that the least
frequently-occurring function of clicks in these data is the one that is most
frequently mentioned in the phonetic literature.
7.5.3 Results: Inter-Speaker Variation
The results of inter-speaker variation in click production are presented
in Table 7.1, looking first at clickers versus non-clickers. The leftmost column in
212
Table 7.1 presents the length of time over which clicks were analyzed, while the
second and third columns represent the numbers of speakers who were found
to be either clickers or non-clickers. A clicker is defined as a speaker who has
been found to click at least once in the given speech sample, and a non-clicker is
defined as a speaker who does not click at all in the given speech sample.
Table 7.1: Number of clickers versus number of non-clickers over varying speech sample
lengths
length clicker non-clicker

1 minutes 39 61
2 minutes 56 44
3 minutes 67 33
4 minutes 72 28
5 minutes 75 25
As seen in Table 7.1, if one considers each sample in its entirety, the proportion
of clickers to non-clickers is around 3:1 (75:25). However, this proportion
could not be arrived at by examining a shorter sample, as the number of non-
clickers decreases as sample length increases, owing to the fact that so many of
the speakers click very infrequently. This can be seen in Figure 7.4, in which it
is apparent that 74% of the DyViS population clicks five times or fewer over the
five-minute period, i.e. they have a click rate of one click per minute or less.
Figure 7.4 displays the number of speakers on the y-axis and number of total
clicks on the x-axis.
213
30
Distribution of the Total Number of Clicks
20
15
10
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 23 24 25 26 27 28 29 53 54
Number of Clicks
Figure 7.4: Distribution of click totals over five minutes of speech
Approximately 50% of speakers click only once, twice or not at all. And while
the mean number of clicks for the group as a whole is 4.26 clicks over five
minutes of net speech, this is highly skewed by three speakers who produce a
very high number of clicks (24, 28, and 54). The mean number of clicks per
speaker drops to 3.4 clicks when the three most extreme clickers are removed.
Figure 7.5 presents the mean click rates in clicks per minute (clicks/min.),
rather than as a cumulative number of clicks, as seen in Figure 7.4. The y-axis
presents the number of speakers that fall within a given range and the x-axis
depicts click rate in clicks per minute.
214
Mean Click Rate
30
20
15
10
3.30-3.45
0.15-0.30
0.30-0.45
0.45-0.60
0.60-0.75
0.75-0.90
0.90-1.05
1.05-1.20
1.20-1.35
1.35-1.50
1.50-1.65
1.65-1.80
1.80-1.95
1.95-2.10
2.10-2.25
2.25-2.40
2.40-2.55
2.55-2.70
2.70-2.85
2.85-3.00
3.00-3.15
3.15-3.30
3.45-3.60

4.8
5.6
0.0-0.15

10.8
Click Rate (clicks/minute)
Figure 7.5: Distribution of click rate (clicks/minute) in DyViS population
There is an inevitable positive skew to the distribution of mean click rate in
Figure 7.5. The mean click rate for the population is 0.88 clicks/min, with a
range of 0.00 clicks/min to 10.8 clicks/min. The standard deviation of the
means is 1.41 clicks/min. There are two suspected outliers at 3.00 clicks/min
and 3.50 clicks/min. There are also three extreme outliers at 4.80 clicks/min,
5.60 clicks/min, and 10.80 clicks/min.
The cumulative distribution graph of mean click rates in Figure 7.6
shows the percentile at which a given click rate falls in relation to the
population. The y-axis represents the cumulative proportion of the population,
and the x-axis presents click rates in clicks per minute.
215
Mean Click Rate
100

90
80
70
60
50
40
30
20
10
0
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
2.4
2.6
2.8
3.0
3.2
3.4
3.6
4.8
5.6

10.8
Figure 7.6: Cumulative percentages for click rate
The curve in Figure 7.6 starts at 25% for those with a click rate of 0, meaning
that 25% of the DyViS population have no clicks present in their speech. From
the first point at 25%, the curve is characterized as having an approximately
logarithmic growth. Figure 7.6 shows that roughly 70% of the population have
click rates at or below 0.8 clicks/min, and only 30% have larger click rates.
7.5.3.1 Discussion
Clicking, as a measure, has been shown to be highly sensitive to sample
length (see Table 7.1), and it is not possible to specify a threshold sample
duration for determining click rate, as the sample duration is dependent upon
frequency of clicking. For example, to determine that someone has a click rate
of, say, 0.2 clicks per minute, it would be necessary to have a sample five
minutes in length, during which time the speaker clicks only once. However, to
216
establish that someone had a click rate of, say 10 per minute, all one would need
is one minute of speech or indeed less. This assumes, of course, that the
clicks would be evenly distributed across time. And, as will be seen in the
section below, such an assumption of intra-speaker stability is not supported by
the data. For the present, however, it is noted that the low number of click
totals for the majority of speakers makes the discrimination capacity of clicks
difficult to establish. Nevertheless, there is potential for clicks to be a good
discriminant for the handful of speakers who produce high click totals, if these
speakers are relatively stable and consistent in their clicking behavior.
7.5.4 Results: Intra-Speaker Variation within an Interaction
The results for intra-speaker variation are presented in Figure 7.7.
Speakers are represented on the x-axis and the click rates (clicks per minute) on
the y-axis. A speakers mean click rate is represented by a black dot and the
vertical bars indicate the range between the minimum and maximum click rate
they attained in any individual minute of speech.
217
Click Rate Ranges by Speaker
16
mean
14
12
10
0
0 20 40 60 80 100
Speakers
Figure 7.7: Mean and range of click rates across all speakers
It is clear from Figure 7.7 that intra-speaker stability generally increases as
mean click rate increases, such that the higher-rate clickers have a greater
range of variability across the individual minute blocks. Thus, even for those
speakers for whom clicks might serve as a potentially discriminant feature, the
clicks tend to occur in localized clusters rather than being evenly spread
throughout the sample. This effectively means that in order to establish that
someone has a high click rate, the analyst would need a relatively large amount
of speech from him/her. In the forensic context, questioned recordings
containing around one minute of net speech from the target speaker are not
unusual. Obtaining five minutes of net speech is much less common. Thus, the
possibility of using clicks as a discriminant feature in forensic casework, even
for high-rate clickers, is quite limited.
218
There is a limited amount of data with which we can calculate variability.
Nevertheless, inter-speaker variation is presented for the DyViS population.
Caution must be exercised when interpreting the SD data. Figure 7.8 presents
the distributions of standard deviations for the population.
Standard Deviation
30
Number of Speakers
25
20
15
10
Standard Deviation (clicks/minute)
Figure 7.8: Distribution of standard deviation in click rate
Figure 7.8 has a positively skewed distribution, like that seen for mean click
rate in Figure 7.5. There are two suspected outliers in the population (at 1.95
clicks/min and 2.07 clicks/min), and one extreme outlier at 5.40 clicks/min.
The mean SD for click rate in the population is 0.69 clicks/min, with a range of 0
clicks/min to 5.40 clicks/min. The SD of the SDs for click rate is 0.70, which is
actually higher than the mean, indicating a large spread in click rate values.
The cumulative distribution graph of SD for click rate is presented in
Figure 7.9. The y-axis shows the cumulative proportion of the population with a
SD at a given point, and the x-axis presents click rate in clicks per minute.
219
Standard Deviations for Click Rate
100
Cumulative Proportion (%) 90

80
70
60
50
40
30
20
10
0
0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 2.25 5.50
Click rate (clicks/minute)
Figure 7.9: Cumulative percentages for standard deviation in click rate
The curve in Figure 7.9 is similar to the curve seen in Figure 7.6, but is slightly
more gradient than logarithmic in its growth. The data in Figure 7.9 show that
25% of the speakers have SDs under 0.25 clicks/min., due to the 25% of
speakers who do not click at all in their five minutes of net speech. The variance
ratio for click rate is 4.06, which signifies that there is more variation between
speakers than within speakers for click rate. A variance ratio of 4.06 is the
highest that has been achieved for any parameter in the current thesis
(articulation rate, long-term fundamental frequency, and long-term formant
distributions). Despite a good variance ratio, caution has to be exercised, as it
must be remembered that there were on average only five click tokens per
speaker.
220
7.5.4.1 Discussion
The sporadic distribution of clicks might be accounted for by the
clustering of click opportunities (i.e. places where clicks can be used as
discourse markers). There is no reason to assume that the need to express the
affective meanings and perform the interaction management functions that
clicks can fulfill should be evenly spread across time. A more detailed analysis
might therefore address the question of the occurrence of clicks as a proportion
of click opportunities. Clustering of click opportunities can, of course, occur
across interactions as well as within them, i.e. some types of conversation may
well present more opportunities than others. For the present, however, another
aspect of intra-speaker variation in clicking is examined, namely possible
accommodation effects.
7.5.5 Results: Intra-Speaker Variation across Different Interactions
Accommodation, the tendency for speakers to adjust their speech
towards that of their interlocutor, has been well documented in respect of a
range of linguistic features (c.f. Giles, 1973; Giles and Ogay, 2007; Shepard, Giles
and LePoire, 2001; Trudgill, 1981).
The click data considered so far were all drawn from the Task 2
recordings of the DyViS database, where each of the 100 subjects conversed
with the same interlocutor, Int1. The recorded interviews that make up the
Task 1 recordings involved two different interlocutors, Int2 and Int3,
conversing with the 100 subjects. The further work reported in this section was
triggered by the informal observation that the subjects appeared to be clicking
more frequently in the Task 1 recordings when speaking with Int2 and Int3
221
than in the Task 2 recordings when speaking with Int1. This observation
provided the motivation to undertake two further analyses: (a) establishing
subjects actual click rates in the Task 1 recordings relative to the Task 2
recordings, and (b) examining the click rates of Int2 and Int3 relative to Int1.
The latter was undertaken with a view to determining whether any increase in
subjects clicking behavior might be accounted for by an interlocutor
accommodation effect.
Fifty subjects were selected at random from the Task 1 recordings, 25
speaking with Int2 and 25 speaking with Int3. As with the Task 2 sampling
procedure, the first two minutes of the conversations were excluded from the
analysis to allow for settling-in time. Three minutes of net speech were
extracted for each subject for comparison with an equivalent three-minute
sample from the Task 2 recordings. Click rates were then compared across the
two tasks. The comparisons showed that, although there was no statistically
significant difference between the numbers of clickers versus non-clickers
(using a chi-squared test, where p = .7401 (Int1 to Int2) and p = .0880 (Int1 to
Int3)), clickers did show a marked increase in click rate when speaking to Int2
and Int3 over when speaking to Int1. The results are summarized in Table 7.2.
The first column identifies the interlocutor, and the second and third columns
present the mean and median click rates, respectively, for the given
interlocutor.
222
Table 7.2: Summary of Speakers Mean and Median Click Rates - Int1 versus Int2 and Int3
interlocutor mean median

Int1 0.72 0.67
Int2 1.60 1.33
Int1 1.53 0.33

Int3 2.16 0.67
Int1* 1.00 0.33

Int3* 1.75 0.67
*denotes rates with outlier excluded
The increase across Int1 to Int2 is significant at the 1% level (using a Wilcoxon
signed rank test, p = .0034 and n = 25). That between Int1 to Int3 falls just short
of significance at this level (1%), but achieves it if one speaker whose high click
rate (speaker 07s click rate is 14.33 clicks/min for Task 1 and 12 clicks/min for
Task 2) is excluded as an outlier (Wilcoxon signed rank test, p = .0076 and n =
24).
The actual changes mean, minimum and maximum - for speakers are
represented in Tables 7.3 and 7.4 (Int1 versus Int2 and Int1 versus Int3). The
first column identifies the direction of change in click rate. The second column
in Table 7.3 and 7.4 identifies the number of speakers with a given change in
click rate, and columns three through five present the mean, minimum, and
maximum changes in click rate for the group of speakers.
223
Table 7.3: Changes in click rate across speaker - Int1 versus Int2
Number of
(Int2-Int1) Mean Minimum Maximum
Speakers
Increase 15 1.601 0.003 3.003
Same 4
Decrease 6 -0.334 -0.003 -0.670
Table 7.4: Changes in click rate across speaker - Int1 versus Int3
Number of
(Int3-Int1) Mean Minimum Maximum
Speakers
Increase 17 1.264 -2.333 4.333
Same 0
Decrease 5 3.444 -0.003 -0.333
In attempting to account for the increases in click rates when subjects spoke to
Int2 and Int3, click rates for Int2 and Int3 were calculated from three randomly-
selected Task 1 recordings. The sampling procedure entailed extracting three
minutes of net speech after the settling-in period, thus providing a total net
sample of nine minutes for each interlocutor. For Int1 an equivalent portion of
post-settling-in speech was extracted from the Task 2 recordings with the same
three subjects selected for Int2 and Int3, thereby providing a total net sample of
18 minutes. The mean click rates for the three interlocutors are set out in Table
7.5. The first column identifies the interlocutor and the second column presents
the mean click rate for the given interlocutor.
Table 7.5: Mean click rates of the three interlocutors
interlocutor mean click rate (clicks/minute)

Int1 1.44
Int2 3.67
Int3 4.56
224
Given the click rates established for subjects from the Task 2 recordings, Int1
might be seen as a relatively average clicker. Int2 and Int3, however, would be
considered relatively high-rate clickers. In view of this, a plausible explanation
for the increased click rates of the subjects when conversing with Int2 and Int3
would be that they are accommodating their clicking behavior towards that of
their interlocutors. It is, of course, entirely possible that the accommodation
effect is bilateral and that interviewers also adjust their click rates towards
those of the subjects. The data to test this view are not available within the
present study, however. Nor is it possible to assess whether interviewer gender
is a factor40; it may or may not be significant that Int1 is a young male, while
Int2 and Int3 are young women. An alternative, or indeed additional,
explanation of the differences might be that the Task 1 interactions offer more
clicking opportunities, these being mock police interviews in which the subjects
are asked questions that might well have them searching for words in
answering. However, this would not account for the relatively high click rates of
Int2 and Int3, and although there are currently no formal findings to present on
this, the clear impression is that there are no obvious differences amongst click
opportunities.
The overriding limitation when analyzing the discriminatory power of
clicks is the unfeasibility to calculate a numerical LR and evaluate the strength
of evidence. The absence of an LR calculation for clicks is due entirely to the fact
that a model does not currently exist with which it might be calculated.
40 Accent may also be a factor, since only one of the interlocutors was also an SSBE speaker.
225
However, there are a number of mathematical procedures that can be used to
arrive at a numerical LR. In forensics the different procedures are selected in
relation to the characteristics of the data distributions. Aitken and Taroni
(2004 p. 37) state that for any particular type of evidence the distribution of
the characteristic [parameter] is important. This is so that it may be possible to
determine the rarity or otherwise of any particular observation. Therefore it is
important to use the model that best fits the distribution of data in order to
represent the strength of evidence as accurately as possible.
Clicks are a particularly complicated form of speech evidence to work
with when used to calculate numerical LRs, as they are discrete in nature. Aside
from DNA profiling (which works with discrete data), there is a lack of methods
when data are discrete rather than continuous. In forensic speech science, there
has not been any LR research that has carried out a comprehensive analysis of
discrete data. LR research in FSS has previously focused on continuous data
(Gold and Hughes, 2013), for which it is possible to assume normality. Once an
assumption of normality is made theory then allows for multivariate
continuous data to be modeled using the means and covariances only (Aitken
and Gold, 2013, p. 148). However, for discrete data a description of the
distribution as normal is not possible.
In this particular case, where there is a desire to calculate LRs for
clicking rate in speakers, there are two main issues to consider when seeking
how to model the data appropriately. The first is the possibility for each
discrete data entry (e.g. the 5-minute recording) to have multiple levels of
response (e.g. a click count for each minute in the recording). For example, in
the present data, multiple levels of response are represented by the multiple
226
click counts over a given amount of time. More specifically, Speaker A may have
5 minutes of net speech, where each minute of net speech yields an individual
count (e.g. 0,0,1,2,0). The second issue is the correlation that exists between
counts. Given that 25% of the population was found not to click at all over 5
minutes of net speech, it is apparent that correlations exist between counts, and
these must be accounted for in a model.
The work in Aitken and Gold (2013) explores the issues and limitations
involved in calculating LRs for discrete data. A Poisson distribution and
bivariate Bernoulli model are proposed for evaluating clicks and any other
discrete data that act in a similar way to clicks. The models proposed in Aitken
and Gold are basic models; however they illustrate issues that need to be
considered in the analysis of discrete data and provide a foundation on which
other models may be built (Aitken and Gold 2013 p. 154). Likelihood ratios
are provided in Aitken and Gold (2013, p. 153). However, they are based on a
limited data set whereby and (set distributions of the population) were not
based on structural learning41 but intuitive guesses about the population
distribution. The LR results for clicks were between 0.30 and 3.35 (i.e. giving
very limited evidence for support) which are small but intuitively sensible
(Aitken and Gold, 2013, p. 154). More practical work is needed to further
develop the models. However, it is hoped that further testing will also produce
smaller LRs. Intuitively, this would align with there being a finite number of
possible clicks produced over the course of a minute, high intra-speaker
variation, and low inter-speaker variation.
41Structural learning makes decisions based on the data at hand, and uses those data to inform
a given model/algorithm/framework (Porwal et al., 2013)
227
Given the limited strength of evidence reported for clicks in Aitken and
Gold (2013), the lack of models for calculating click-rate LRs may not be all that
devastating. This is due to the general lack of capacity of click rate to
discriminate between speakers of English. As always, exceptions are to be
made, however, for those individuals who lie towards the margins of the
distribution curve and who can be classified as outliers with respect to click
rate.
7.7 Conclusion
While it would be dangerous to generalize beyond the variety of English
analyzed in this study, the view of those forensic practitioners surveyed in
Chapter 3 who considered tongue clicking to be a highly discriminant feature of
speaker behavior is largely unsupported by the present data for young male
speakers of SSBE. Firstly, there is insufficient variation across the majority of
speakers analyzed for the variable to provide a reliable index of speaker
individuality. Secondly, even for the high-rate clickers who stand apart from
the majority, there is within-conversation instability to the extent that one
would need speech samples of a length seldom encountered in questioned
forensic recordings in order to reliably establish an overall click rate. Thirdly,
intra-speaker variation also occurs across interactions, apparently as a result of
accommodation towards the clicking behavior of interlocutors. This suggests
that rate of clicking rather than being solely a property of an individuals
speech production practices, might usefully be viewed as resulting from an
interaction between speaker and interlocutor. The question remains, then, of
whether it is worth considering clicking at all when conducting speaker
228
comparison casework. In spite of these findings, it is suggested that, in certain
cases, it may well be. Studies such as those of Wright (2007; 2011a; 2011b) and
Ogden (2013) on the interactional functions of clicks, as well as the general
observations of phoneticians on the functions of clicks in conveying attitudinal
and affective meanings, provide normative data and descriptions. For this
reason, these studies allow forensic practitioners to assess the speech samples
they examine for the occurrence of non-normative, i.e. idiosyncratic, usage.
Such occurrences may be of assistance in the comparison task, and in this
respect forensic phoneticians are indebted to their non-forensic counterparts
for providing valuable resources. This is, in fact, just a further instance of a
more general indebtedness of the forensic speech community to work in
mainstream academic research in linguistics and phonetics. As noted in French
and Stevens (2013), sociophoneticians and dialectologists have provided
normative descriptions of language varieties that serve as backcloths for the
evaluation of findings in speaker comparison cases.
Unless it were to transpire that patterns of clicking behavior are
different for other varieties of English or (for example) differ in accordance with
speaker age or gender - and nothing has been found in the sociolinguistic
literature on English to support that view - the mere comparison of click rates
across samples is in the overwhelming majority of cases unlikely to advance the
speaker comparison task, for the reasons outlined above.
229
Chapter 8 Overall Likelihood Ratios
8.1 Introduction
The whole is greater than the sum of the parts. - Aristotle
The survey of FSC practices (Chapter 3) revealed that for the vast
majority of expert forensic phoneticians, it is the overall combination of
parameters that they consider crucial in discriminating between speakers
(despite some parameters having greater weight than others). For this reason,
the current chapter addresses the issue of combining parameters for speaker
discrimination through empirical testing.
The combination of phonetic, linguistic, and non-linguistic parameters in
an FSC has traditionally been carried out by experts through implicit mental
calculations. That is to say, an expert creates a mental representation of the
properties of an individuals speech and makes a judgment about the likelihood
that the speakers in the suspect and criminal samples are the same person
(based on the combined weight of the evidence). The process by which an
expert mentally combines parameters to arrive at a conclusion is not
transparent. As such, it has been argued that different experts will weigh certain
parameters more highly than others, based purely on personal opinions (Rose
and Morrison, 2009). For this reason, the traditional method of parameter
combination in FSCs is highly subjective and is difficult to replicate.
Bayes theorem on the other hand offers a more explicit and transparent
alternative for the combination of parameters. A simple combination procedure,
known as nave Bayes (Kononenko 1990) involves multiplying the individual
230
LRs (or equivalently, the addition of individual LLRs) assuming that there is
mutual independence between parameters (i.e. the parameters are not
correlated). The combination of correlated parameters is a problem for the LR
framework, because unless parameters are mutually independent there is a risk
of over-estimating the strength of evidence by considering the same parameter
more than once. Alternative methods such as logistic-regression fusion, MVKD,
and Bayesian Networks have been put forward to circumvent the problem
whilst maintaining a Bayesian approach (Aitken and Lucy, 2004; Brmmer et
al., 2007; Gonzalez-Rodriguez et al., 2007; Morrison et al., 2010). However,
given the lack of appropriate testing, it is unclear whether logistic-regression
fusion or MVKD adequately take account of correlations in the data.
The aim of the present chapter is to amalgamate the individual speech
parameters from Chapters 4-7 into a complete system42, whereby discriminant
power, strength of evidence, and validity can be tested for all analyzed speech
parameters in combination. Previous research has developed methods to
facilitate the combination of individual speech parameters into some form of a
combined system. However, the blend of approaches taken in this chapter has
never been used before. The chapter begins by exploring the existing
relationships between LTFD, AR, F0, and click rate to check for potential
correlations. The correlation coefficients are then used to inform appropriate
combination methods given the (in)dependencies that exist amongst the given
42 A system is defined by the Oxford English Dictionary (http://www.oxforddictionaries.com/

definition/english/system) as a set of things working together as parts of a mechanism or an
interconnecting network; a complex whole. The term system is often used in the ASR
literature to refer to a complex whole that is well-suited to providing a response to two
competing hypotheses being tested (the evidence given the prosecutors hypothesis divided by
the evidence given the defenses hypothesis). The term system is extended in this thesis to both
individual speech parameters and speech parameters in combination that can also provide the
basis for an evaluation of the two competing hypotheses.
231
parameters. After the combination of individual parameters into a complete
working system, the discriminant ability, strength of evidence, and validity are
tested for the combined parameters. The integration of methodological
approaches employed in this chapter is a first for calculating overall likelihood
ratios (OLRs). This approach is intended to demonstrate how an analyst would
go about using these methods in order to avoid an over- or under-estimation of
the strength of evidence for calculating OLRs for the data under scrutiny.
The combination of forensic speech evidence under a numerical LR framework
has received a reasonable amount of attention in the literature (Alderman,
2004; Kinoshita, 2002; Rose et al., 2003; Rose et al., 2004; Rose and Winter,
2010). The main focus of this earlier research is the issue of combining
parameters that are potentially correlated. Early studies evaluating traditional
linguistic phonetic parameters often recognized this problem but did nothing to
try to ameliorate it. For example, Kinoshita (2002) used nave Bayes to combine
LRs based on the best-performing set of formant predictors from /m/, //, and a
set of short vowels into a single expression of posterior probability. Similarly,
Alderman (2004) generated an OLR from different vowel formant predictors
using nave Bayes in order to compare the speaker-discriminatory performance
of different combinations of parameters (and individual features of
parameters).
Rose et al. (2003) displayed a more overt awareness of the issues
surrounding correlation within and between parameters. In their study, they
compared the discriminatory performance of formants using segmental cepstra
232
from /: / in Japanese. Linear regression was applied to the parameters to
assess the degree of correlation only within individual parameters (i.e. the
formants). The individual LRs were combined into an OLR using an assumption
of independence, although between-parameter correlation was never explicitly
tested because it was assumed that given the very different phonetic nature of
the three segments used, there was unlikely to be much correlation between all
but their highest formants (Rose et al. 2003 p. 195). Rose et al. (2003) make a
good attempt at accounting for within-parameter correlation, but fail to go one
step further to test the between-parameter correlations. Phonetic theory would
predict that the parameters are not correlated. However, without further
testing, correlations may go unexposed (and unrealized). Rose et al. (2003) also
note that linguistic theory leads them to believe that the higher formants may
be correlated, yet nothing is done to account for it. Therefore, it is probable that
the results produced for the study were over- or under-estimations of the
The development of LR modeling techniques has brought with it the
capability of dealing more appropriately with the complexities of correlation.
Aitken and Lucys (2004) MVKD formula treats the set of data from which LRs
are computed as multivariate data, and as such is able to account for within-
segment correlation. Rose et al. (2004) investigated the comparative
performance of the multivariate LR approach and the nave Bayes assumption
of independence. The nave Bayes approach was shown to overestimate the
strength of SS and DS LRs compared with the more conservative MVKD model.
The proportion of errors was also better when independence was assumed,
which led Rose et al. (2004) to conclude that the correct formula is still not
233
exploiting all the discriminability in the speech data and (as such) the Idiots
approach nave Bayes is still preferable (Rose et al. 2004 p. 496). However
the study fails to discuss the fact that nave Bayes produces misrepresentative
estimates of the strength of evidence when parameters are correlated, which
could lead to a miscarriage of justice in a real case.
Another recently-adopted technique to account for potential correlation
between phonetic-linguistic parameters in LR-based FSC is the logistic
regression fusion approach. Fusion is a form of back-end processing (Rose
and Winter, 2010, p. 42) which attaches weights to parameters based on
correlations between LRs from individual parameters. This contrasts with
front-end processing which considers correlations in the raw data. Fusion
was developed within the field of ASR (Brmmer et al., 2007; Gonzalez-
Rodriguez et al., 2007; Ramos Castro, 2007) and has since been applied in a
number of studies using traditional phonetic parameters (Morrison, 2009;
Morrison et al., 2010; Rose, 2010b; Rose, 2011) leading Rose and Winter (2010,
p. 42) to claim that fusion is one of the main advances to have emerged from
automatic methods.
Fusion is currently the only alternative to a nave Bayes approach for LR-
based forensic phonetic analysis. However, there are a number of potential
problems with fusion. Firstly, back-end processing, as the name suggests, deals
with correlations after the generation of numerical LRs has been performed.
Therefore, as suggested by Rose, it is possible that two segments which
are not correlated by virtue of their internal structure and which therefore
should be naively combined nevertheless have LRs which do correlate (Rose
2010, p. 32). Equally, the reverse is possible, whereby correlated parameters
234
generate non-correlated LRs. More broadly, there is also an issue of efficiency.
Since fusion is implemented after the generation of LRs, the original analysis
may unnecessarily include a number of highly-correlated parameters, which
when combined provide a limited strength of evidence.
8.3 Data
The present study does not introduce any new data, as it works with the data
presented in Chapters 4-7. The parameters under consideration are mean long-
term formant frequency distributions (for F1-4), mean articulation rate, long-
term mean fundamental frequency, and click rate.
8.4 Correlations
This section considers potential correlations that exist within and between
parameters. Correlations are calculated for the speakers as a group, as opposed
to individual speakers. Therefore, it could be the case that the correlations
found for the group of 100 speakers do not exhibit the same patterns as those
calculated for an individual speaker.
8.4.1 Methodology
Correlations were calculated to identify potential relationships or
mutual independencies within and between parameters. Two groups of
correlations were calculated for the data: those within LTFD (i.e. LTFD1, LTFD2,
LTFD3, LTFD4) and those between parameters (LTFD1-4, AR, F0, click rate).
The formants within LTFD are treated as individual parameters for correlation
testing, given that phonetic theory has established that individual formant
235
measurements represent different physiological aspects of a vowel (e.g.
frontness/backness of the tongue, height of the tongue, voice quality;
Ladefoged, 2006; Laver, 1994). Correlation coefficients were calculated for
individual LTFD comparisons by selecting two LTFD measurements at a given
data point, resulting in hundreds of data points per formant comparison.
Calculating correlation coefficients between parameters required a single data
point per person so a mean value was calculated for each speakers LTFD1-4,
AR, and F0 (i.e. three separate means). The data for click rate already existed as
a single data point for each speaker, so no additional mean calculations were
required.
All correlations in this section were calculated using Spearmans rank
correlation coefficient. This method was preferred over Pearson correlation
coefficients, as the latter assesses how well the relationship between two
variables can be described using a monotonic function. The Pearson correlation
coefficient is calculated on the assumption that the relationship between two
variables is linear. Because the pair-wise relationships between variables under
consideration are not known (nor can they be assumed to be linear), the
Spearmans rank correlation coefficient was the logical choice.
MatLab (version R2012a) was used to create scatterplots and to
calculate the correlation coefficients for all pairs of parameters. Table 8.1
presents all six possible pairing combinations for the LTFD parameter, and
Table 8.2 presents all 15 possible pairing combinations between parameters.
236
Table 8.1: Formant pairings within LTFD
Parameter 1 Parameter 2
Long-term Formant Distributions
LTFD1 LTFD2
LTFD1 LTFD3
LTFD1 LTFD4
LTFD2 LTFD3
LTFD2 LTFD4
LTFD3 LTFD4
Table 8.2: Between-parameter pairings
Parameter 1 Parameter 2
LTFD1 Mean F0
LTFD1 AR
LTFD1 Click Rate
LTFD2 Mean F0
LTFD2 AR
LTFD2 Click Rate
LTFD3 Mean F0
LTFD3 AR
LTFD3 Click Rate
LTFD4 Mean F0
LTFD4 AR
LTFD4 Click Rate
Mean F0 AR
Mean F0 Click Rate
AR Click Rate
Tables 8.1 and 8.2 are complete lists of all 21 parameter pairings. The first and
second columns simply identify the parameters that are being compared against
each other.
The point at which two parameters can be deemed to be correlated is a
matter of subjective judgment, in that there is no specific correlation coefficient
that explicitly signifies dependence between two parameters. The decision of
237
independence is made by the expert, which in turn can result in different
opinions regarding the threshold at which correlations are implied (ultimately,
this can cause variation in the LR results). For the purpose of this study,
correlations were considered through structural learning (see 7.6), which is
informed by the data rather than theoretical considerations. Final correlation
judgments were made by me after examining scatterplots in conjunction with
correlation coefficients for each pair-wise comparison. My judgments relating to
correlations were also confirmed by a forensic statistician (Marjan Sjerps, p.c.).
8.4.2 Within-Parameter Correlation Results
The scatterplots for all pair-wise comparisons within LTFD are
presented in Figures 8.1 - 8.6. The y-axis presents one LTFD parameter, while
the x-axis represents another.
r = -0.16
LTFD1 (Hz)
Figure 8.1: LTFD1 versus LTFD2
238
r = 0.05
LTFD1 (Hz)
LTF
r = -0.03
Figure x:
LTFD1 (Hz)
239
r = 0.43
LTFD2 (Hz)
r = 0.20
LTFD2 (Hz)
240
r = 0.41
LTFD3 (Hz)
The scatterplots in Figures 8.1 - 8.3 do not exhibit any strong relationships
between the variables, and graphically suggest that there are no correlations.
Figures 8.4 and 8.6 are characterized as having moderate positive correlations,
while Figure 8.5 has a slightly weaker positive correlation. The correlation
present in Figure 8.5 (LTFD2 vs. LTFD4) is most likely representative of indirect
correlation, given that LTFD2 correlates with LTFD 3, and LTFD3 correlates
with LTFD4.
The correlation coefficients for within-LTFD comparisons are presented
in Table 8.3. The intersection of a column and row indicates a given comparison,
and the value within the box is Spearmans rank correlation coefficient (r). A
value closer to 1 or -1 suggests that two parameters are correlated, while a
value close to 0 suggests the two parameters are not correlated.
241
Table 8.3: Correlation coefficients within LTFD
LTFD1 LTFD2 LTFD3 LTFD4

LTFD1 1.00 -0.16 0.05 -0.03
LTFD2 1.00 0.43 0.20
LTFD3 1.00 0.41
LTFD4 1.00
Based on the results seen in Figure 8.1 - 8.6 and Table 8.3, an informed
judgment can be made with regard to which parameters appear to be correlated
within LTFD1-4. The results suggest that LTFD2 is correlated with LTFD3,
LTFD3 is correlated with LTFD4, and LTFD2 is indirectly correlated with LTFD4
(they have a transitive relationship by way of LTFD3; this also referred to as a
partial correlation). LTFD1 and LTFD2 have a correlation coefficient of -0.16.
However, this correlation was not deemed to be significant (r is less than 0.25;
confirmation also given by Marjan Sjerps, p.c.). Therefore, LTFD1 is reasoned to
be independent from LTFD2-4.
8.4.3 Between-Parameter Correlation Results
The scatterplots for all pair-wise comparisons between parameters are
presented in Figures 8.7 - 8.12. The y-axis represents the first parameter, and
the x-axis represents the second parameter.
242
r = -0.13
r = 0.15
Mean LTFD1 Mean LTFD2
r = -0.14 r = -0.12
Figure 8.7: Mean AR versus LTFD1-4
243
r = -0.04
AR (syllables/second)
Figure 8.8: Click rate versus mean AR
r = -0.03
AR
F0 (Hz)
Figure 8.9: Click rate versus mean F0
244
r = 0.10 r = -0.22
r = -0.13 r = -0.20
Figure 8.10: Click rate versus LTFD1-4
245
r = 0.08
r = 0.19
r = -0.06
r = -0.07
Figure 8.11: Mean F0 versus LTFD1-4
246
r = -0.06
AR (syllables/second)
Figure 8.12: Mean F0 versus mean AR
The scatterplots in Figures 8.7 - 8.12 do not exhibit any signs of strong (or even
moderate) correlations between any of the parameter pairings. Because the
scatterplots have a limited number of data points (only 100 in this case)
compared with the number of data points for Figures 8.1 - 8.6, the calculation of
correlation coefficients is necessary (as was also seen in Table 8.3) to quantify
the levels of correlation between parameters.
The results of the correlation coefficients are presented in Table 8.4.
Table 8.4: Correlation coefficients within- and between-parameters
LTFD1 LTFD2 LTFD3 LTFD4 AR F0 Click

Rate
LTFD1 1.00 -0.16 0.05 -0.03 0.15 0.08 0.10
LTFD2 1.00 0.43 0.20 -0.13 0.19 -0.22
LTFD3 1.00 0.41 -0.14 -0.06 -0.13
LTFD4 1.00 -0.12 -0.07 -0.20
AR 1.00 -0.06 -0.04
F0 1.00 -0.03
Click Rate 1.00
247
In terms of between-parameter correlations, Table 8.4 does not present any
new strong correlations (those within LTFD have already been discussed in
8.4.2). The strongest relationships found between parameters are those for
LTFD2 vs. click rate (-0.22), LTFD4 vs. click rate (-0.20), and LTFD2 vs. mean F0
(0.19). Linguistics literature and phonetic theory do not give any reason to lead
one to believe that these parameters should be related to one another43.
Therefore, the very weak correlations seen in Table 8.4 have most likely
happened by chance. Correlation does not imply causality, and these three cases
appear to be good examples of this.
8.4.3.1 Discussion
Based on the results seen in Figures 8.7-8.12 and Table 8.4, an informed
judgment can be made with respect to the parameter correlations for the data
set. The results suggest that there is no parameter correlation between LTFD,
mean AR, mean F0, and click rate (confirmation given by Marjan Sjerps, p.c.). As
such, these parameters are deemed to be mutually independent from one
another for this particular data set.
8.5 Overall Likelihood Ratios
Based on the interdependencies and conditional dependencies found in 8.4,
OLRs can be calculated for the system. A model does not currently exist with
43 However, it is possible that F0 and F2 could be related. High F2 values are associated with
tongue fronting, and as the tongue body fronts, it pulls on the hyoid bone, from which the larynx
is suspended. Laryngeal tension of this sort would promote higher F0, because of tension on the
vocal folds. Therefore, it is possible that one might anticipate a correlation between high F2
values and high F0 values.
248
which calculate numerical LRs for click rate, and therefore the OLRs calculated
in this section exclude this parameter from analysis.
8.5.1 Methodology
The OLRs presented in this section were calculated in multiple stages.
Individual LRs were calculated for LTFD1, AR, and F0 using a MatLab
implementation of Aitken and Lucys (2004) MVKD formula (Morrison 2007),
and a separate LR was calculated for LTFD2-4 together using the same MVKD
formula. This was done in order for the algorithm to take into account the
correlations that exist between these three parameters (LTFD2-4). An intrinsic
methodology, whereby the test and the reference speakers came from the same
population of 100 speakers, was used for all LR calculations. Speakers 1-50
were used as the test speakers, while speakers 51-100 served as the reference
speakers.
The results from the individual LRs and the LR from LTFD2-4 were then
multiplied together following Nave Bayes (given that 8.4 demonstrated that
AR, F0, LTFD1, and LTFD2-4 were independent of one another) to form a
complete system. Additional variations of the system were also computed in the
same manner as for the complete system, whereby the LR for LTFD2-4 is always
calculated together (in the MVKD formula) and multiplied by the other
individual LRs in different combinations. A MatLab script44 was then used to
calculate basic statistics, EER, and Cllr for the OLR system and variations on this
system.
44 This script was developed by Phil Harrison of J P French Associates.
249
Logistic-regression calibration was also applied to the complete system
in two different orders using a MatLab script45. Logistic-regression calibration
(see 8.5.3 for further discussion) was applied in the first instance to individual
parameters before combination, and applied in the second instance to the
complete system after the parameters had been combined in order to compare
the effectiveness of the calibration (in terms of EER and Cllr).
8.5.2 Overall Likelihood Ratio Results: Uncalibrated
The results of the OLR for the complete system are provided in Table 8.5.
The complete system is composed of LTFD1, LTFD2-4, F0 (mean and standard
deviation), and AR (mean). The first column in Table 8.5 presents the
comparison type (SS or DS pairs), followed by the percentage of correct pairs,
mean LLR, minimum LLR, and max LLR. The final two columns report on the
complete systems validity where the sixth column provides the EER and the
final column presents Cllr.
Table 8.5: Summary of LR-based discrimination for the complete system (100 speakers)
Complete System SS 92.00 5.673 -3.082 7.316
.0607 .3793
Complete System DS 93.27 1.560 -infinity 3.963
Table 8.5 shows that the combination of all parameters into the complete
system provides an EER of 0.0607, and a Cllr of 0.3793. It appears that the
complete system is good at identifying SS pairs, and slightly better at identifying
DS pairs. The strength of evidence that the system offers is considerably
stronger than that seen in the tests reported in Chapters 4-6. Figure 8.13
45This script was created by Niko Brmmer, modified by Geoffrey Morrison, and edited by
Vincent Hughes.
250
presents the Tippett plot of the complete system, and Figure 8.14 is a zoomed-in
version of Figure 8.13.

Figure 8.13: Tippett plot of the complete system
251
Figure 8.14: Zoomed-in Tippett plot of the complete system
Figures 8.13 and 8.14 illustrate the distribution of SS and DS pairs. The strength
of evidence of the DS pairs is higher than the strength of evidence offered by the
SS pairs. Following Champod and Evett (2000), the system has the potential to
offer strength of evidence (either for the prosecution or defense hypotheses)
that is considered very strong support. Figure 8.14 shows that the crossover
between the curves representing the comparison of SS and DS pairs is very
close to the zero threshold, but not on it, and it is possible that calibration of the
system might improve its validity (see 8.5.3 for analysis).
Although the complete system in Table 8.5 includes all available
parameters it is necessary to consider the possible performance of other
252
combined systems should they outperform the complete system. Table 8.6
provides ten alternative systems to the complete system from Table 8.5. The
organization of Table 8.6 follows that of Table 8.5.
Table 8.6: Summary of LR-based discrimination for alternative systems (100 speakers)
LTFD1+LTFD2-4+F0 SS 92.00 5.691 -3.517 7.348
.0631 .4322
LTFD1+LTFD2-4+F0 DS 92.82 1.528 -infinity 4.151
LTFD1+LTFD2-4+AR SS 98.00 3.811 -1.423 5.380
.1310 .6101
LTFD1+LTFD2-4+AR DS 73.06 1.311 -infinity 4.082
LTFD1+LTFD2-4 SS 94.00 3.807 -1.618 5.426
.1361 .6348
LTFD1+LTFD2-4 DS 71.43 1.387 -infinity 4.508
LTFD1+F0+AR SS 86.00 2.432 -3.900 3.594
.0709 .4780
LTFD1+F0+AR DS 95.43 0.134 -infinity 2.373
LTFD1+F0 SS 82.00 2.150 -4.335 3.353
.0771 .5266
LTFD1+F0 DS 94.41 0.096 -infinity 2.306
LTFD1+AR SS 76.00 1.046 -1.967 2.143
.2284 .7873
LTFD1+AR DS 77.14 0.083 -infinity 2.101
LTFD2-4+F0+AR SS 96.00 5.220 -2.149 6.887
.0647 .4160
LTFD2-4+F0+AR DS 89.22 1.469 -infinity 3.848
LTFD2-4+F0 SS 96.00 5.249 -2.585 6.933
.0707 .4742
LTFD2-4+F0 DS 88.20 1.465 -infinity 3.817
LTFD2-4+AR SS 100.00 3.319 0.457 4.951
.0929 .8413
LTFD2-4+AR DS 60.20 0.789 -infinity 3.213
F0+AR SS 88.00 1.625 -2.959 2.457
.0855 .4197
F0+AR DS 91.47 0.048 -268.938 2.143
None of the alternative systems in Table 8.6 outperforms the complete system
in terms of validity. The next best performing system in terms of EER (after the
complete system) is that of LTFD1+LTFD2-4+F0 with an EER of 0.0631 and Cllr
of 0.4322. This second-best system is identical to the complete system minus
the inclusion of AR, which suggests that the inclusion of more parameters
improves the systems validity.
253
8.5.3 Overall Likelihood Ratio Results: Calibrated
Calibration is a procedure for improving a systems precision whereby a
well-calibrated system is considered to be more reliable (DeGroot and Fienberg,
1983). Calibration was first utilized by weather forecasters (DeGroot and
Fienberg, 1983), but has since made its way into automatic speaker comparison
(Ramos-Castro et al., 2006), and phonetic/linguistic-based FSCs (Morrison,
2012). Ramos-Castro et al. (2006, p. 6) have shown the importance of the
calibration of LR values computed by an automatic system, arguing that highly
discriminant likelihood ratios might achieve a high performance in terms of
probability of error of the posterior probabilities. However, a high calibration
loss46 in the computed LR values may lead to arbitrarily high errors. For this
reason, logistic-regression calibration (using a cross-validation method) has
been applied here to the complete system in two different orders to compare
calibrated results. Figures 8.15 and 8.16 illustrate the first method, in which
parameters were calibrated individually and then combined. Figure 8.17
illustrates the results of the second method, where individual parameters were
combined and the complete system was then calibrated.
46 Quantified according to the degree to which LR values incorrectly support a hypothesis.
254
Figure 8.15: Tippett plot of the complete system - parameters calibrated individually and then
combined
255
Figure 8.16: Zoomed-in Tippett plot of the complete system - parameters calibrated
individually and then combined (-6 to 6 LLR)
256
Figure 8.17: Zoomed-in Tippett plot of the complete system- system calibration after
combination of parameters (-10 to 10 LLR)
The calibration of individual parameters before combination (see Figures 8.15
and 8.16) resulted in an EER of .0554 and a Cllr of .2831. There was an
improvement in both the EER and Cllr from the uncalibrated system of .0053
and .0962, respectively. The calibration of the complete system after the
combination of parameters in Figure 8.17 resulted in an increase (i.e. a higher
value) of EER of .0011, and an improvement (i.e. a lower value) of Cllr of .1408.
Results show that for the complete system, calibration before combination
provides the best EER, while calibration after combination provides the best
Cllr. The differences between the two methods are minimal. However, one
improves the gradient result for incorrect/correct judgments (Cllr) while the
other improves the hard detection error rate (EER). In forensic speaker
257
comparison, a protocol for the order in which the application of calibration
should take place has not been previously discussed. Therefore, more research
is needed on the effects of the order of operations in which calibration is
applied. It is also entirely possible that calibration could be applied twice, once
before combination and once after. However, that was not tested here.
8.6 Discussion
This section focuses on three key discussion points: whether clicks can help
improve the performance of the complete system, the comparison of results
from all systems combinde to the results found for the individual parameters,
and potential limitations of a combined system.
8.6.1 Do Clicks Improve the Complete System?
The uncalibrated complete system achieved an EER of .0607, while
correctly identifying 92% of SS pairs and 93.27% DS pairs. The complete
system, however, did not include click rate as one of the combined parameters,
as click rate did not lend itself to the calculation of numerical likelihood ratios.
The system should now consider whether click rate has the potential to help in
discriminating between the SS and DS pairs that were judged incorrectly.
A total of 165 DS pairs (out of a possible 2450 DS pairs) were judged
incorrectly by the complete system. Four SS pairs were judged incorrectly.
When those 169 incorrectly-judged pairs are extracted, it is possible to identify
those pairs that include any of the speakers that had extreme outlying click
rates (see 7.5.3). The three extreme outliers in terms of click rate (speakers
007, 024, and 033) form one half of the pairings in 20 DS pairs. If those extreme
258
click rates were considered within the complete system, one could propose that
those 20 DS pairs would then be judged correctly. This would then increase the
percentage of correct DS pairs to 94.08% (an additional increase of 0.81%). The
extreme click rate outliers are not a part of any of the incorrectly-judged SS
pairs, so click rate will not help further the comparison in these pairings.
The small increase in correctly-identified DS pairs would be unlikely to
improve the EER or Cllr dramatically. However, it is possible that it would
improve the system performance to some degree. Although click rate could be
used to help discriminate those SS and DS pairs that were incorrectly identified,
there is the potential for click rate to also decrease the system performance. If
click rate were to make the correctly-identified SS pairs significantly more
dissimilar and the correctly-identified DS pairs significantly more similar, the
performance of the complete system would be decreased further. However, it is
important to note that this discussion of system performance (where click rate
is included) remains hypothetical without including all of the click rate data. As
we saw in Chapter 7, click rate appears to be highly variable within and
between speakers, which in turn characterizes click rate as an unstable
parameter (to a higher extent than AR, even). The inclusion of this unstable
parameter could cause more variation in OLRs, which could in turn weaken the
system. It is also important to consider that Aitken and Gold (2013) showed
that the LRs produced for click rate were associated with relatively weak
strength of evidence. Therefore, perhaps the inclusion of click rate in all 2500 SS
and DS comparisons may not contribute significantly to the OLRs, leaving the
performance of the complete system relatively unchanged.
259
8.6.2 Comparing Individual Parameters to the Systems
The uncalibrated complete system in 8.6.2 performed relatively well in
respect of the systems validity with a number of the alternative systems
performances following closely behind. The extent of their achievements is best
shown in juxtaposition with the performance of the individual parameters prior
to their being placed into a combined system. Table 8.7 contains all the
individual parameters performances. The organization of Table 8.7 is identical
to that of Table 8.5 and 8.6.
Table 8.7: Summary of LR-based discrimination for individual parameters (100 speakers)
LTFD1 SS 72.00 0.224 -2.158 1.902
.2806 .8840
LTFD1 DS 71.70 -4.858 -68.768 1.993
LTFD234 SS 74.00 0.649 -8.461 4.996
.0798 .9023
LTFD234 DS 95.31 -25.812 -166.915 3.046
F0 SS 92.00 0.958 -3.404 1.936
.0849 .4547
F0 DS 89.90 -24.204 -269.159 1.906
AR SS 90.00 0.180 -1.480 2.060
.3340 .8981
AR DS 46.20 -2.940 -8.760 0.820
Table 8.7 shows that LTFD234 has the lowest EER at .0798, followed by F0 at
.0849. LTFD1 and AR both have EERs around .30. The best-performing
individual parameter in terms of Cllr is F0 at .4547, with the remaining three
parameters close to .90. The results for the individual parameters would
suggest that a system including LTFD234 and F0 will have the best opportunity
of performing well in terms of system validity. It also appears that a system that
includes AR will benefit from SS comparisons, but will potentially be weakened
by its DS comparisons.
260
Table 8.8 provides a summary of the improvements and deteriorations
in the complete system performance compared to the best-performing
individual parameter for a given LR statistic. The first column identifies the LR
statistic, and the second column indicates whether or not the complete system
improved, deteriorated, or stayed the same in comparison to the best-
performing individual parameter (i.e. LTFD1-4, AR, or F0). The final column
indicates the degree of the change identified in the second column.
Table 8.8: Performance comparison between individual parameters and the complete system
Results Change Degree of Change

EER Improve .0191
Cllr Improve .0754
SS % Correct Same 0.00
DS % Correct Deteriorate 2.04%
SS Mean LLR Improve 4.715
DS Mean LLR Deteriorate 27.380
SS Min LLR Deteriorate 1.602
DS Min LLR Improve Infinity
SS Max LLR Improve 2.32
DS Max LLR Deteriorate 3.143
Table 8.8 shows that the complete system outperformed any individual
parameter in terms of EER, Cllr, SS Mean LLR, DS Min LLR, and SS Max LLR
(highlighted in light blue). The complete system deteriorated in performance
with respect to DS % Correct, DS Mean LLR, SS Min LLR, and DS Max LLR
(highlighted in dark blue). There was no change observed between the
complete system and best-performing individual parameter in the performance
of SS % Correct (highlighted in mid-blue). Overall, Table 8.8 shows that the
complete system has (most importantly) the best system validity as well as the
most improvement in terms of strength of evidence for SS. The strength of
261
evidence for a single parameters tends to be very limited for SS (see Table 8.7),
despite the DS strength of evidence being relatively strong. Therefore an
improvement in strength of evidence for SS comparisons is highly desired.
The same LR statistics presented in Table 8.8 are also considered with
respect to the best overall performing system or individual parameter in Table
8.9. Table 8.9 presents the LR statistic in the first column and identifies which
system or individual parameter performed the best in that respect.
Table 8.9: Best-performing system or individual parameter in relation to LR statistics
Results Best System/Individual Parameter

EER Complete System
Cllr Complete System
SS % Correct LTFD234+AR System
DS % Correct LTFD1+F0+AR System
SS Mean LTFD1+LTFD234+F0 System
DS Mean LTFD234
SS Min LTFD234+AR System
DS Min Complete System
SS Max LTFD1+LTFD234+F0 System
DS Max AR
Table 8.9 identifies whether the complete system (light blue), an alternative
system (mid-blue), or an individual parameter (dark blue) performed best for
the given LR statistic. The complete system remains the best in terms of system
validity (which is the most important of the statistics). The alternative systems
achieve the best performance for five of the LR statistics, while individual
parameters are the best performing for two of the LR statistics. The results in
Table 8.9 confirm the opinions set out by experts in 3.10 that the inclusion of
more parameters results in better overall speaker discrimination (i.e. EER).
However, the level at which more is better is not all-encompassing with
262
respect to all the LR statistics. It is the case that there are smaller systems and
two individual parameters that outperform the complete system in some
respects. Therefore, the general opinion of experts - that more parameters are
better for discriminating between speakers - needs to be redefined insofar as it
does not appear to be the case that the more parameters that are used for
speaker discrimination, the better the system validity.
8.6.3 Limitations
Despite the relatively good performance of the complete system in terms
of validity it still has four important limitations to consider. The first limitation
is that the complete system is not outperforming alternative systems and
individual parameters in relation to all of the LR statistics. As discussed in
8.7.2, the inclusion of more parameters does not necessarily correspond to
better performance in all respects. Results suggest that the addition of the
right parameters could increase performance but the addition of poorly-
performing individual parameters may not improve the overall system. This
poses a dilemma regarding which parameters to include in the FSC analysis.
Should the expert only select the best-performing parameters, in terms of EER,
or should he/she try to include all parameters that characterize an individuals
speech? Additionally, the expert must recognize that with the addition of
parameters comes the (potential) additional uncertainty introduced in the
system. This issue also needs to be addressed in respect of some type of
confidence interval, and it may be the case that the confidence intervals
measure of credibility will be what sets the complete system apart from an
alternative system or individual parameter.
263
The second limitation to the complete system lies in the steps taken
before the combination of parameters, in which dependencies are tested
between and within parameters. In general, pairs of parameters that had
correlation coefficients of less than 0.25 were considered to be independent of
one another. It is possible that the complete system is limited in respect of the
combination of parameters that exhibit some small levels of correlation that go
unaccounted for. An example is the nave LR calculation approach taken in
Chapter 5 for LTFD1-4. MVKD was used to calculate LRs for LTFD1-4. However,
the current chapter found LTFD1 to be independent of LTFD234. This would
dictate treating LTFD1 separately from LTFD234 (as was seen in 8.6.2), and
the LRs from the two sets to be multiplied following nave Bayes. The EER for
the MVKD combination was 0.0414, while the current chapter reports an EER
for LTFD1+LTFD234 of 0.1361. The two methods for combining LTFD1-4 lead
to a dramatic difference in EER. It appears that perhaps not taking small
correlations into account when working with the given data has caused EER to
increase.
The third limitation is the application of calibration. There is no set
protocol for when calibration is to be applied. In 8.6.3 it was demonstrated
that different results can be achieved when the individual parameters are
calibrated separately and then combined (yielding an EER of 0.0618), compared
to the combination of parameters followed by the application of calibration
(giving an EER of 0.0554). One could plausibly consider the calibration of
parameters separately before combination, and the calibration of the system in
a second phase after the combination of parameters has been carried out. More
264
research is needed in order to make an educated decision on the order in which
combination and calibration are performed.
The final limitation is very basic, but it is perhaps the most important. It
concerns the threshold at which two parameters are deemed to be dependent
on one another. A threshold of 0.25 was selected, but this was somewhat
arbitrary. A better understanding is needed through a combination of empirical
testing and theory to allow for more reliable decisions to be made on the
(in)dependence of parameters. To some extent this is being explored in the
International Association of Forensic Phonetics and Acoustics-funded grant
entitled Identifying correlations between speech parameters for forensic
speaker comparisons (Gold and Hughes, 2013). However, further research is
still needed on the relationships between speech parameters in other accents
and languages.
8.7 Conclusion
The results of this study have shown that the combination of parameters into a
complete system improves system performance in terms of validity (EER and
Cllr). It is not necessarily the case that more parameters will improve all aspects
of the system, but where it matters most - in terms of validity - the addition of
more parameters prevails. The combination of the parameters central to this
thesis (AR, LTFD, F0, and to some extent clicks) raises the question of what will
happen when other parameters are added to the system. Following expert
opinion in this respect (see Chapter 3), one would expect validity to further
improve. However, it could be the case that there will be a threshold at which
the addition of parameters can no longer improve validity. It is also possible
265
that the addition of unstable (highly variable) parameters will make the
performance of the system deteriorate. Extrapolating from the current results it
is expected that the strength of evidence for SS pairs will only increase as
parameters are added, while the strength of evidence will remain similar for DS
pairs (as it is already very strong).
It is difficult to predict the performance of a system that includes
additional parameters that may exhibit different variation characteristics from
the current parameters. It is also difficult to extrapolate the performance of the
system while considering parameters (here, click rate) which cannot be
incorporated into a numerical LR. For this reason, the current complete system
can only serve as a building block contributing towards a larger system that will
incorporate a much wider range of linguistic and phonetic parameters, and
which will possibly improve the discrimination level of speakers in FSCs.
266
Chapter 9 Discussion
In this chapter, the results from the previous six chapters (3-8) are summarized
and discussed. The results of the first international survey of FSC practices are
evaluated with respect to four of the valued phonetic and linguistic parameters
selected by expert forensic phoneticians. The four parameters are then
considered individually with regards to their speaker discriminant ability,
strength of evidence, and validity. Finally, the combination of the four
parameters into a human-based speaker comparison system is discussed and
compared with those used in ASR analysis.
9.1 Summary of the Forensic Speaker Comparison Practices

Survey
The results of the first international survey of forensic speaker comparison
practices showed a fundamental lack of consensus on the methods employed in
FSCs. Although the finding might come as a surprise to phoneticians and
linguists working outside FSS, the degree of variation in methods will not be
surprising to those working in various other fields of forensic science (see the
journals Science and Justice or Forensic Science International for a plethora of
articles debating forensic methodologies). Most importantly, the survey gave an
insight into which parameters experts identified as being the most helpful
speaker discriminant parameters above all others. The following are the top five
ranked parameters (in order):
1. Voice quality
2. Dialect/ accent variants and vowel formants
3. Speaking tempo and fundamental frequency
267
4. Rhythm
5. Lexical/grammatical choices, vowel and consonant realizations,
phonological processes, and fluency
The majority of experts indicated that despite some individual
parameters being good speaker discriminants, it is the combination of
parameters that hold the most discriminant power in FSCs. The discriminant
potential of speech parameters in combination rather than on their own is not
often addressed in the research literature.
9.2 Summary of Phonetic/Linguistic and Forensic Findings for

Individual Parameters
Four parameters identified by survey participants as having a high discriminant
value were investigated, namely articulation rate (AR), long-term formant
frequencies (LTFD), fundamental frequency (F0), and clicks. This involved
assessments of the individual parameters as speaker discriminants (percentage
of correctly-classified SS and DS pairs, strength of evidence, and validity), how
well expert expectations of the parameters matched the results, and whether
the results were similar to those reported in previous studies.
9.2.1 Articulation Rate
Speaking tempo, and particularly AR, was identified by 20% of experts in
Chapter 3 as one of the most helpful speaker discriminants (ranked 3rd in 9.1).
The high expectations surrounding the discriminant capacity of AR motivated
empirical testing. Three key observations can be made in relation to the
influence methodology has on the calculation of AR: (1) the definition of the
speech interval does not significantly affect results, (2) varying the minimum
number of syllables in a speech interval does not make AR significantly more
268
stable, and (3) testing suggests that the exclusion of speech segments, and
perhaps the definition of the syllable (i.e. phonetic versus phonological) may
have more effect on ARs than other factors.
The findings tell a different story from that predicted by expert opinion,
suggesting that there are misconceptions about the discriminant capacity of AR.
Under an LR framework, SS pairs were correctly identified 90% of the time,
while DS pairs were correctly identified at a rate less than chance (46.2%).
Articulation rate contributed weak strength of evidence for SS pairs, and only
moderate strength of evidence for DS pairs. AR had an EER of 0.3340 (the
highest EER of the three parameters tested under the LR framework, i.e. AR,
LTFD, and F0) and a Cllr of 0.8981. AR as an individual speaker discriminant
was found to be rather weak. A simple impressionistic determination of
speaking tempo, rather than a tedious and potentially unnecessary quantitative
analysis of AR, may be sufficient in most forensic cases. Despite apparent
misconceptions about the discriminant power of AR, it should nevertheless
remain a tool in a forensic phoneticians toolbox as there will always be the
possibility of outlying speakers for which AR may be extremely valuable.
9.2.2 Long-Term Formant Distributions
Vowel formants, including long-term formant distributions (LTFD), were
identified by 28% of experts (Chapter 3) as being among the most useful
speaker discriminants (ranked 2nd in 9.1). This provided the motivation for
further discriminant testing. The results from the analysis of LTFD provide both
phonetically- and forensically-relevant results. In terms of the phonetic findings,
there are two pertinent observations in relation to methodology and speaker
269
specificity: (1) small changes in the package length for LTFD have only a small
effect on results, and (2) higher formants (LTFD3 and LTFD4) are suggested to
carry a greater amount of speaker-specific information than lower ones.
The forensic findings confirm experts expectations regarding the
discriminant potential of vowel formants. Under an LR framework, the
combination of LTFD1-4 correctly identified SS pairs 84% of the time, while DS
pairs were correctly identified 97.4% of the time. As a system, LTFD1-4 had an
EER of 0.0414 (the lowest EER of the three parameters tested under the LR
framework: AR, LTFD, and F0) and a Cllr of 0.5411. Despite the promising
findings of LTFD1-4 as a combined system, 8.4.2 unexpectedly revealed that
LTFD1 was statistically independent of LTFD2-4, and should technically be
treated separately (as an independent parameter). If LTFD1 is treated
separately, the combined LTFD2-4 system still achieves a low EER of .0798 and
a Cllr of 0.9023, where SS pairs and DS pairs are correctly identified 74% and
95.3% of the time, respectively. These findings, in combination with previous
findings from Becker at al. (2008), Moos (2010), French et al. (2012), and Jessen
et al. (2013), suggest that LTFDs perform very similarly to MFCCs under
comparable data conditions, and, as an individual speaker discriminant, LTFD is
rather strong. The only potential limitation of LTFD is that it averages across all
vowels, which in turn eliminates idiosyncrasies and habituations of certain
vowels that relate accent information. Unless a single vowel phoneme can yield
more promising results, the evidence suggests that LTFD should be considered
over individual vowel analysis under the LR framework47.
Including both could be seen as doubling evidence, insofar as LTFD measurements encompass the
47
multiple formant measurements made for individual vowel phonemes.
270
9.2.3 Long-Term Fundamental Frequency
Fundamental frequency (F0) was identified by 20% of experts (Chapter
3) as being one of the most helpful speaker discriminants (ranked 3rd in 9.1).
This motivated empirical testing of the parameter. The results from the analysis
of F0 provide both phonetically- and forensically-relevant results. In terms of
the phonetic findings, there are two main observations. Firstly, small changes to
the package length of F0 only have a small effect on the results (as we saw with
LTFD). Secondly, it was reported in 8.4.3 that, unexpectedly, F0 did not
correlate with LTFD1-4. Given previous research (Narang et al., 2012; Syrdal
and Steele, 1985) it might have been expected that F0 and LTFD1 would be
correlated, especially since using Lombard speech it has been shown that F1
increases as F0 increases (Kirchhbel, 2010). The independence of F0 and
LTFD1 was also reported by Moos (2010), and may be an indication that F0 and
F1 correlations can only be found when vowels are analyzed individually as
phonemes (Narang et al., 2012; Syrdal and Steele, 1985). However, once F0 is
compared to an LTFD that relationship is lost, perhaps because (i) F0 and F1 are
not correlated for all phonemes (and an averaging of phonemes eliminates any
strong correlation present in the data), or more likely (ii) there is non-vowel
information included in the acoustic signal which suppresses any correlation
that might be present.
The forensic findings confirm experts general expectations regarding the
discriminant power of F0. Under an LR framework, F0 correctly identified SS
pairs 92% of the time, and DS pairs 89.9% of the time. F0 contributed a rather
weak strength of evidence for SS pairs, while DS pairs had a much stronger
strength of evidence. As a system, F0 had an EER of 0.0849 (the second highest
271
of the three parameters tested under the LR framework: AR, LTFD, and F0) and
a Cllr of 0.4547. Given the findings and the plethora of previous LR research on
F0, it is suggested that F0 as an individual speaker discriminant is rather strong
within a contemporaneous recording. However, it is not entirely clear how well
F0 can discriminate between individuals when same-speaker evidence comes
from non-contemporaneous recordings. Previous literature would suggest that
F0s discriminant power will decrease when same-speaker evidence from
different recordings is introduced in addition to any deletrious (external)
factors (e.g. disguise, recording transmission, vocal effort; see 2.2 for more
factors). The study by Boss (1996) gave an example of F0 mismatch in a real
forensic case. The difference between the F0 in the criminal and suspect
recording was 88Hz, due in large part to situational differences in the
recordings (the suspect sounded more nervous in the criminal recording than
the suspect recording (Boss, 1996, p. 156)). Unless it were to transpire that F0 is
robust to many of the factors detailed in 2.2, the mere comparison of mean F0s
and SDs is on its own unlikely to advance the speaker comparison task
dramatically. However, as always exceptions are to be made for those
individuals who can be classed as outliers, and using F0 in conjunction with
other speech parameters for FSCs is suggested.
9.2.4 Click Rate
Non-linguistic parameters (which include clicks) were identified by 18%
of experts (Chapter 3) as being amongst the most useful speaker discriminants.
Again, this motivated empirical testing. The results from the analysis of click
frequency provide both phonetically- and forensically-relevant results. In terms
272
of the phonetic findings, there are two pertinent observations: (1) discourse
analysis classifications can lend themselves to the quantification and
categorization of speech parameters (Wright, 2005), and (2) accommodation
effects are present in clicks, in that the rate of clicking, rather than being solely a
property of an individuals speech production practices might arise from an
interaction between speaker and interlocutor.
The forensic findings suggest that there are misconceptions surrounding
the discriminant capacity of clicks. While it would be dangerous to generalize
beyond the variety of English analyzed in this thesis, the view of those forensic
practitioners surveyed in Chapter 3 who considered tongue clicking to be a
highly discriminant feature of speaker behavior is largely unsupported by the
present data for young male speakers of SSBE. Click data is positively skewed
and discrete, and there is currently no method available for deriving an LR from
them (although see Aitken and Gold (2013) for current developments in
proposed algorithms for calculating the LRs of clicks). For this reason, the
discriminant capacity can only be assessed qualitatively and with reference to
the population statistics. Further, there is insufficient variation across the
majority of speakers analyzed for the variable to provide a reliable index of
speaker individuality. Additionally, even for the high-rate clickers who stand
apart from the majority, there is within-conversation instability to the extent
that one would need speech samples of a length seldom encountered in criminal
forensic recordings in order to reliably establish an overall click rate. Given the
high degree of intra-speaker variation and restricted inter-speaker variation,
clicking frequency is a rather weak parameter. Unless it were to transpire that
patterns of clicking behavior are different for other varieties of English or differ
273
in accordance with speaker age or gender - and nothing has been found in the
sociolinguistic literature on English to support that view - the mere comparison
of click rates across samples is, in the overwhelming majority of cases, unlikely
to advance the speaker comparison task. However, as with AR, it is advised that
clicks remain a tool in an experts toolbox for those speakers identified as
outliers.
9.3 Summary of Discrimination Performance by the Overall

System
This section has outlined the main forensic findings for articulation rate (AR),
long term formant frequencies (LTFD), fundamental frequency (F0), and clicks
as a combined, overall system. The forensic findings are assessed with respect
to the overall systems success at discriminating between speakers (percentage
of SS and DS pairs correct, strength of evidence, and validity), how well expert
expectations of the parameters corresponded with the results, and if the results
are similar to those found in previous studies.
A large majority of the experts discussed in Chapter 3 indicated that it is
the combination of speech parameters that makes for better performance at
speaker discrimination than individual parameters. Therefore, the combination
of the parameters in this thesis was motivated by expert opinion and the lack of
human-based systems that test speaker discrimination of parameters in
combination. Under an LR framework, the complete system (LTFD1, LTFD2-4,
AR, F0) correctly identified SS pairs 92% of the time, and DS pairs 93.3% of the
time. The combined system contributed very good strength of evidence for SS
pairs, and even stronger strength of evidence for DS pairs. Overall, the combined
274
system had an EER of 0.0607 and a Cllr of 0.3793. After the complete system
was calibrated, the EER and Cllr decreased to 0.0554 and 0.2831, respectively.
The combination of parameters into a complete system therefore improves
system performance (in comparison to individual parameters) in terms of
validity (EER and Cllr). It is not necessarily the case that the inclusion of more
parameters (e.g. SS % correct, DS % correct, mean SS LLR) improves all aspects
of the system, but where it matters most (validity) the addition of parameters
does result in an improvement. The results of the complete system are almost as
good as those from ASRs under similar conditions (French et al., 2012). Table
9.1 compares the results of the research presented in this thesis against those
from the ASR in French et al. (2012).
Table 9.1: Human-based results against ASR (Batvox) results from French et al. (2012) on
studio quality data
Same Speaker Different Speaker
Current study 92% 94.1%
French et al. (2012) 100% 95%
The percentages presented in Table 9.1 indicate the proportion of SS and DS
comparisons judged correctly where studio-quality recordings were used. Given
that the system developed in this thesis only incorporates three parameters
under an LR framework, the incorporation of more speech parameters might
improve system performance further.
Out of all the views advanced by experts that were reported in Chapter 3,
perhaps the most valuable and the most accurate is that the combination of
parameters results in the best speaker discrimination performances. The
275
current system serves as a starting point from which to expand. Additional
research needs to be carried out on the discriminant power of parameters in
combination.
9.4 Overall Findings

There are two principal findings that emerge from the present research: (1) the
performance of the human-based system created in this thesis is comparable to
ASR performances on the same studio-quality recordings, and (2) the FSS
community is faced with many obstacles if they wish to continue to align
themselves with other, more developed, forensic sciences by implementing an
LR framework for FSCs.
9.4.1 Human-Based System versus ASR
The research conducted for this thesis was not intended to provide a
comparison of the efficacy of human-based systems and their ASR counterparts.
Nevertheless, through the development of this human-based system over the
course of the current research project, the methodologies employed have made
quantitative comparisons between the human-based system and ASRs possible.
ASRs used for FSCs are typically known for demonstrating their validity through
error rates, and the testing of these ASRs is easily replicable. ASR error rates are
typically presented in terms of the frequencies of false negatives and false
positives. A false negative is the classification of a target trial as a non-target
trial, and a false positive is a non-target trial classified as a target trial (van
Leeuwen and Brmmer, 2007). The human-based methodology (using phonetic
and linguistic parameters) for FSCs has become known for not providing error
rates and for a lack of replicability. However, it is shown that by adopting a

276
numerical LR framework, a human-based system can also provide validated
results, while fostering tests that are easily replicable.
The performance of the human-based system, consisting of LTFD, AR, F0,
and click rate, is comparable to that of an ASR tested on the same type of data
(high-quality studio recordings). The human-based system created for this
thesis reported false positive errors (different-speaker comparisons = LR > 1) of
5.9%48 and false negative errors (same-speaker comparisons = LR < 1) of
8.0%49. An ASR system tested on the same data (French and Harrison, 2010)
reported false positive errors of 4.5%, and achieved zero false negatives.
It is important to note here that the performance of a human-based
system is dependent upon the expertise of the analyst. It is likely that some
degree of cross-analyst variation would be observed. For the present human-
based system, LTFD and F0 would be least susceptible to inter-analyst variation
given that the methodology for extracting data is relatively automatic50. AR and
click rate calculations are more dependent on the analyst. For example, when
calculating AR the analyst must decide which speech to include in an interval
and what to ignore, and for clicks the analyst must decide whether s/he is
hearing a click rather than a percussive. For this reason, caution should be
exercised so as not to overestimate the replicability of results.
9.4.1.1 A Fair Comparison?
The comparison of the human-based system and an ASR investigated by
French et al. (2012) showed that the ASR only minimally outperformed the
48 This includes click rate.

49 This does not include click rate.
50 See Jessen and Becker, 2010 (discussed in 5.2) and Konrat and Jessen, 2013 for variation in
LTFD and F0 results when measured by multiple analysts.
277
human-based system on studio-quality data. However, such a comparison puts
the human-based system at a considerable disadvantage. ASRs have consistently
been shown to perform well under high-quality data conditions (Campbell,
1997; French and Harrison, 2010; Reynolds, 2002; Reynolds et al., 2000). The
ASR, however, is susceptible to degradation in system performance as the data
quality gets worse. French and Harrison (2010) tested an ASR on real case data
where the outcomes of the cases were known to the authors (i.e. guilty/not
guilty verdicts). For the 767 comparisons undertaken, the ASR achieved an EER
of 24.2%. When the ASR was tested on the cases in which the system judged the
technical quality of the data to be adequate, it accepted 171 of the 767
comparisons and produced an EER of 5.4%. When the ASR was further tested on
recordings that the system judged to be only marginally adequate, an EER of
15.1% was returned for the 369 comparisons. The results indicate a steep fall in
performance for the ASR when processing less than ideal quality recordings.
A human-based system may be able to offer a better performance than
ASRs when testing lower-quality recordings. For example, the 767 real forensic
comparisons used for testing with an ASR by French and Harrison, 2010 had
previously been analyzed by the authors using a phonetically- and linguistically-
based methodology. None of those 767 comparisons resulted in a known
miscarriage of justice, and for all comparisons their conclusions were in
agreement with those made by the trier of fact.
In real forensic cases where recordings are of poor quality or short
duration, in conclusion, a human analyst may be better equipped to extract data
and make conclusions than an ASR.
278
9.4.1.2 Scope for Improvements in the Human-Based System
The human-based system created for this thesis was limited to four
parameters owing to the inherent time restrictions of the research. Given the
analysis presented in 8.6 and 8.7, it is reasonable to assume that the addition
of good speaker discriminants would increase the validity of the system. With
enough additional parameters it is possible that the human-based system could
eventually outperform an ASR on high-quality studio data. The addition of
parameters should not be done ad hoc, but should involve phonetically- and
linguistically-informed choices. This would involve: (i) selecting parameters
that are good speaker discriminants (e.g. voice quality, VOT, or those
parameters reported by experts in 3.9 or 3.10,), and (ii) not selecting
parameters that are significantly correlated with others (but if they are,
weighting the correlations appropriately).
Additionally, it is appropriate to consider the integration of this human-
based system with ASRs. This could potentially be similar to the Vocalise 51
software package created by Oxford Wave Research Ltd (Vocalise, 2013) that
measures MFCCs in addition to traditional phonetic parameters (e.g. F0 and
LTFD; see 10.2 for further discussion of Vocalise). Given the good performance
of the ASR on the data tested, and the good performance achieved by the
human-based system, a combination of the two could result in an even better
overall performance. If integration was to be done it would probably be best to
choose between the inclusion of MFCCs or LTFDs, as they effectively analyze the
same aspects of a speaker (i.e. vocal tract resonances). LTFD analysis provides
the analyst with an indication of the speakers habitual use of the vowel space,
51 http://www.oxfordwave research.com/j2/products/vocalise [Accessed 8 August 2013]
279
which is strongly correlated with the dimensions of the vocal tract (Moos, 2010;
French et al., 2012). MFCC analysis reports on approximately the same aspect,
as MFCCs are essentially abstract representations of the dimensions of the vocal
tract. For this reason, LTFDs and MFCCs are highly correlated (French et al.,
2012) and the inclusion of both parameters would result in the doubling of
evidence, which in turn could lead to a miscarriage of justice. Therefore, given
that MFCCs (marginally) outperform LTFDs (as shown in 5.2), it is suggested
that MFCCs should be selected over LTFDs. An integrated system could then
consist of ASR analysis, F0, AR, and click rate, which again is entirely possible if
ASR (MFCC) data are found to be independent of the other parameters included
in the integrated system. Without testing this integrated system it is difficult to
say with certainty that discriminant performance would improve. However,
with the exchange of LTFD for MFCC that is likely to be so.
9.4.1.3 The Trade-Off
When comparing the performance of the human-based system with
ASRs, it is important to consider a trade-off that occurs when human
intervention is involved. The human intervention is the analysis and
examination undertaken by an analyst for a given comparison (e.g. through
sound file editing, time taken for an analysis). Irrespective of methodology
(human or ASR) the level of human intervention typically needs to be
increased52 as the quality of the recordings (or duration) being compared
decreases. In turn, the level of human intervention utilized in a FSC creates a
52In ASR analysis more time would be required of the analyst for the editing of recordings
(cleaning them up, for example, with a bandpass filter), and for human-based analysis more
time would be required for the analysts to extract measurements and additional data from the
poor quality recordings.
280
trade-off with the variation that may be present in the FSC results. The more
human intervention required, the more cross-analyst variability may be present
in results. For example, if an FSC comparison was vital to a criminal case, rather
than having an ASR completely reject a case that was of poor recording quality
(or offer a conclusion with a very high EER associated with it), a human-based
analysis could be attempted in an effort to extract any relevant conclusion.
9.4.2 Obstacles Facing the Implementation of an LR Framework
The exercise of creating a human-based system for this thesis has
revealed a number of difficulties that surround the LR framework and its
application to FSCs. Those shortcomings are as follows:
1. Subjective elements of the methodological process

2. Delimiting the relevant population
3. Availability of population statistics
4. Lack of models available to calculate LRs
5. Appropriate combination of parameters
The LR framework is intended to create a separation between an experts bias
and the facts of the evidence; however, within FSCs the application of the LR
framework still has elements of subjectivity. It is possible to alter methodologies
(e.g. package lengths of LTFD or F0, as seen in 5-6) in order to achieve a more
desirable strength of evidence. For this reason, it is argued that the LR
framework is not completely objective. However, the levels of subjectivity are
far less under an LR framework than with other frequentist frameworks.
The present study did not address the debate surrounding the
delimitation of the relevant population, as intrinsic LRs were carried out on the
known optimal population (i.e. the test and reference sets of speakers came
from the same linguistically homogeneous corpus). However, the delimitation of
281
the relevant population will remain a difficulty facing FSCs, as was also argued
in French et al. (2010). For this reason, a mutually agreed protocol or
methodology would need to be proposed in order for the continuing debate to
subside. See Hughes (in progress) for further information and discussion on the
issue of the relevant population.
The lack of population statistics, as previously mentioned by French and
Harrison (2007) and French et al. (2010), is a very real problem for the
implementation of a numerical LR. The time, effort, and resources needed to
collect a sufficient quantity of population statistics are almost limitless. As a
testament to this, for the present study it took nearly three years to collect
population statistics for just four parameters for 100 speakers in a single and
very specific population (and, moreover, these data were extracted from a
previously-collected database). All things being equal, following a conservative
number of possible parameters to analyze in a FSC of around 60 (just for
example), the collection of sufficient population statistics for a delimited
population would take approximately 45 years53 to complete by a single
person. At that point, given the occurrence of sound change, it would be time to
scrap the collected population statistics and start again. Such a feat hardly
seems practical. Therefore, in order to continue the development of the LR
framework for FSCs, alternatives need to be put in place.
The lack of appropriate models is currently one of the biggest challenges
faced by the FSS community, because as it stands only certain parameters can
be included in a numerical LR framework. A FSC does not simply consist of a few
vowel formant measurements, and phoneticians would argue that it takes a
53 45 years = 60 parameters at 4 parameters collected every 3 years
282
number of parameters beside vowels to play a role in characterizing an
individuals speech. For this reason, more linguistically-motivated models are
required to enable the incorporation of previously unrepresented
phonetic/linguistic parameters in an LR framework (see Aitken and Gold, 2013
for the proposal of a new, linguistically-motivated model).
Finally, further research should focus on the identification and testing of
appropriate methods for combining speech evidence. There are currently
multiple options for this, but little testing has been done to compare such
methods. It is also the case that only three of these methods (i.e. fusion, MVKD,
GMM, and nave Bayes) are ever really implemented in FSC. If the numerical LR
framework is to become the way of the future, then research should consider
the use of Bayesian Networks for combining speech evidence; these have
already been successfully implemented in more developed forensic disciplines
(e.g. DNA; Evett et al., 2002).
9.5 Methodological Limitations

As with most empirical research, methodological shortcomings are often
inevitable, and the research presented in this thesis is no exception. This section
outlines and discusses three general methodological limitations: (1) the absence
of non-contemporaneous data, (2) the use of intrinsic LR testing, and (3) using
only speakers 1-50 for all the same-speaker comparisons when calculating LRs.
The recordings used in the present study were all obtained from single
recording sessions, as there were no non-contemporaneous data available for
any of the 100 speakers. Non-contemporaneous speech samples have been a
frequent topic of discussion in previous research, with an increasing number of
283
studies citing the importance of incorporating recordings that are made several
days, weeks, months, or even years apart (see Enzinger and Morrison, 2012;
Loakes, 2006; Morrison et al., 2012b; Nolan et al., 2009; Rhodes, 2013).
However, I would argue that there are a number of other external factors that
have more of an impact on results than the separation of recordings by a mere
few days (such as the effects of accommodation reported in 7.5.5). In any case,
the results presented here, despite the fact that they are based on the use of
contemporaneous recordings, nevertheless provide a general baseline in
relation to the discriminant abilities and practicalities of a human-based system.
The second methodological shortcoming is the absence of extrinsic
criminal samples on which to test the human-based system. The LRs calculated
in the present study were based on the DyViS data set of recordings of 100
speakers, whereby the first 50 speakers always acted as the test samples and
the second set of 50 speakers always acted as the reference samples. This means
that the 100 speakers were in some way a part of either the test or reference
sample, and that tests were not conducted using outside data sets. As a result,
the testing is susceptible to over- or under-estimation of the strength of
evidence as proposed in Rose et al. (2006c, p. 329). However, if an additional
relevant database existed, extrinsic testing would be possible, as the population
statistics for DyViS are now available.
The final limitation is the way the data (i.e. speakers) were divided for
calculating LRs. It is probable that assigning different speakers to act as either
the suspect/criminal or background population would cause variation in the
resulting LRs. For empirical testing purposes, the calculation of intrinsic LRs
typically requires a set of speakers to act as both the criminal and the suspect,
284
while an additional set of speakers must act as the reference population. The
number of speakers chosen to act as the criminal/suspect or the reference
population can vary and is entirely at the analysts discretion. For convenience
this thesis divided the data set evenly. As such, the calculation of LRs produced
50 SS comparisons and 2450 DS comparisons. However, it is entirely feasible for
the speakers to have been divided in a number of alternative ways, which might
have produced different outcomes.
9.6 Implications for Forensic Speaker Comparisons

For forensic phoneticians, the population statistics presented in Chapters 4-7
may serve as a helpful tool for casework. It is also suggested that all forensic
phoneticians should consider the inclusion of LTFDs in their analysis. The
relative ease of extracting LTFDs from speech recordings means that this
parameter could be easily utilized and potentially offers a large contribution to
FSCs.
As things currently stand, not all speech parameters can be incorporated
into a numerical LR. Therefore, the use of a complete 54 numerical LR is
impossible. For this reason, experts are faced with a number of decisions,
should they choose to continue to develop methodologies in an effort to align
with other forensic disciplines. Judge Hodgson from Australia argues that not all
types of evidence can be sensibly assigned an LR, and that therefore there is no
way of mathematically combining all evidence under a Bayesian framework
(Hodgson, 2002). Should the field of FSS agree on this statement, it is worth
considering whether all speech evidence can be sensibly assigned a numerical
54 By complete it is meant that all possible analyzable speech parameters are included.
285
LR. If all speech evidence can be forced into a numerical LR model then an
experts job is done. However, this does not seem a plausible possibility, given
the complexity of speech evidence. For this reason, it is my view that experts are
left with three options for the future: they must take the first two if they wish to
align with other forensic disciplines, and the third, a default option, if they are
content with the status quo:
1. Adopt a verbal form of the likelihood ratio framework

2. Present evidence in the form of numerical LRs (for those parameters
that readily lend themselves to such a framework) and present the
remaining speech parameters using a different conclusion
framework.
3. Continue presenting evidence as they do currently.
Although practitioners of forensic phonetics have accepted that the LR is the
logically and legally correct framework within which to present evidence, the
practicalities of the framework will inevitably contribute to the direction taken
by the majority of experts in the future. A large role in the adoption of an LR
framework will also be played by regulations under the country of practice and
simply practicality issues. All countries and institutions are constrained by the
laws, rules, and regulations in which they work. In 3.11, an expert from China
reported that s/he was required to use a binary conclusion framework by
his/her government employer. If forensic phoneticians have the luxury of being
able to choose their conclusion framework, their decision will come down to a
practicality factor. Given the results of this thesis, I believe that the implications
for the field of FSCs are such that a complete numerical LR is unrealistic, and
that alternatives should be explored (such as 1 and 2 presented above).
286
Chapter 10 Summary & Conclusion
The following chapter reviews the empirical results and discussion detailed in
Chapters 3-9 by relating them back to the four research questions introduced in
Chapter 2. The chapter concludes with suggestions for future work that would
expand upon the research aims of this thesis.
10.1 Research Questions Revisited

phoneticians (around the world) typically analyze in a FSC case and which
parameters do they recommend as being highly discriminant?
Chapter 3 provided insight into parameters commonly used in FSCs. Those
reported parameters were as follows: vowels (formants), consonants (timing
aspects, frequencies of energy loci, auditory quality), F0 (mean, median, mode,
SD, range, alternative baseline), voice quality, intonation, speech tempo (AR and
SR), rhythm, linguistic features (aspects of turn-taking, patterns of code
switching, discourse markers (including clicks), telephone opening and closing
behaviors, lexico-grammatical usage), and non-linguistic features (filled pauses,
audible breathing, laughter, throat clearing).
Out of all the possible parameters analyzed in FSCs, the experts top
ranked parameters (the first three) in terms of the expected discriminant ability
were: (1) voice quality, (2) dialect/accent variants and vowel formants, and (3)
speaking tempo and F0. Three of these parameters (LTFD, AR, F0) and clicks
were chosen for analysis in subsequent chapters.
287
(2) If experts are to provide their opinion on the most helpful speaker
discriminants, will these selected parameters be good speaker
discriminants?
Chapters 4-7 presented the results of the discriminant ability of AR, LTFD, F0,
and clicks. LTFD and F0 showed promising results, with EERs of less than .1. AR
and clicks were not as good at discriminating between speakers. AR had an EER
of .33, and although LRs were not calculated for clicks, the results suggested that
click frequency was a very unstable parameter and unlikely to be a useful
discriminant for the majority of speakers.
a. Do experts expectations surrounding the discriminant value of

certain speech parameters match the parameters empirically tested
performance?
The results presented in Chapters 4-7, in combination with the results from the
survey in Chapter 3, suggest that disparities exist between expert opinions and
empirical findings with regard to AR and clicks. However, expert expectations
appear to be accurate for LTFD and, to a lesser extent, F0.
(3) How well do speech parameters work in combination to discriminate

between speakers?
Table 8.5: Summary of LR-based discrimination for the complete system (100 speakers)
Complete System SS 92.00 5.673 -3.082 7.316
.0607 .3793
Complete System DS 93.27 1.560 -infinity 3.963
Table 8.5 (from Chapter 8) presents a summary of the uncalibrated LR results
for LTFD, F0, and AR in combination. The system combining these three
288
parameters produces a lower EER and Cllr than any individual parameter. The
results supported the view expressed by survey participants, i.e. that the
combination of parameters is more successful in discriminating between
speakers than any of the parameters individually.
a. What steps need to be taken in order to appropriately combine

speech parameters?
The proper combination of speech evidence will vary depending on the given
data set. However, for all cases, correlations should be tested in order to
establish whether there are dependent relationships amongst speech
parameters in order to avoid miscarriages of justice (through the doubling of
evidence). For the given data set, dependent relationships were identified
amongst LTFD2-4, and MVKD was used to account for the existing correlations
(by applying statistical weightings). The remaining parameters were found to be
mutually exclusive and the speech evidence was therefore combined using
Nave Bayes.
b. Are multiple speech parameters in combination always better than

individual parameters at discriminating between speakers? Are more
parameters better?
Chapter 8 revealed that for the given data set, the addition of more parameters
improved system validity. However, this did not improve all of the LR statistics
(e.g. exceptions included SS and DS percent correct, SS and DS Mean LLR, and SS
and DS Max LLR).
(4) What are the practical limitations/implications of using the numerical

LR framework in FSCs?
289
Chapter 9 discussed the limitations and implications of using a numerical LR
framework that arose during the development of the system created for this
thesis. Those limitations include, but are not confined to: subjective elements of
the methodological process, difficulties in delimiting the relevant population,
the limited availability of population statistics, the lack of models to calculate
LRs for complex speech data distributions, and the range of methods used to
combine speech evidence.
a. What recommendations, if any, can be provided by attempting to

implement a numerical LR framework?
Chapter 9 concluded that two fundamental factors the amount of time needed
to collect enough data to create a human-based system consisting of just four
parameters, and the inherent complexity of speech evidence - inhibit the
implementation of a numerical LR framework. Although it has been done, it is
difficult to implement a completely numerical LR framework if one is to do it
properly/responsibly. Chapter 9 therefore recommended that practitioners
wishing to use a Bayesian framework should consider adopting a verbal LR
framework or a combination of a verbal and numerical LR framework, instead
of a purely numerical one.
b. What can a human-based (acoustic-phonetic) system tell the field

regarding the ease with which a numerical LR can be computed for FSCs?
The algorithms with which a human-based system can calculate a numerical LR
are the same as those used by an ASR. However, the time needed to collect and
analyze the data for the actual LR calculation is much more intensive for a
human-based system. Time constraints aside, Chapter 9 suggested that the
290
human-based system was capable of producing results comparable to those of
ASRs. It also proposed that with real forensic material (e.g. degraded or shorter
criminal recordings), the human-based system could potentially outperform an
ASR.
10.2 Opportunities for Future Research
Successful application of the numerical LR framework will require a large
amount of additional research if the presentation of FSC conclusions in such a
framework is to become an everyday reality. The work presented in this thesis
would benefit from future investigations in two prominent areas of research:
(1) the integration of ASRs and phonetic-linguistic parameters, and (2) research
into more transparent and successful ways for combining correlated speech
evidence (e.g. fusion).
As discussed in 9.4.1.2, there is potential for improvements in speaker
discrimination through the integration of ASR techniques and phonetic-
linguistic parameters. Future research in this area may benefit from the
availability of systems such as Vocalise (Vocalise, 2013) which allows for
(semi-)automatic analysis and comparison of samples using LTFDs as well as
MFCCs.
The second challenge how to combine correlated speech parameters in
a transparent and appropriate manner could be explored through the use of
Bayesian Networks. Other, more developed, forensic disciplines (e.g. DNA
analysis) rely on Bayesian Networks (Aitken and Taroni, 2004) as an explicit
and logical method for combining evidence (see Evett et al., 2002). In forensics,
the use of Bayesian Networks for the derivation of FSC LRs enables any
291
correlations that exist between (or within) parameters to be weighted
appropriately. A Bayesian Network therefore avoids over- or under-estimation
of the evidence, by accounting for the correlations through a front-end
processing technique. An example of a Bayesian Network (using the
parameters analyzed in this thesis) is provided in Appendix B.
10.3 Conclusion
This thesis set out to explore viability of aligning the field of FSC with other,
more developed, forensic sciences that are currently implementing a numerical
LR framework. The research has highlighted a number of difficulties that face
the FSS community if experts are to continue in their efforts to align themselves
with advanced forensic disciplines. At the present time, the application of a
completely numerical LR framework is fundamentally impractical, insofar as
the numerical LR is unable to incorporate all pieces of speech evidence that
count (those which are discrete rather than continuous, and which cannot be
adequately quantified). In the process of addressing the main aims of this thesis,
additional findings were presented, including: the survey of FSC practices, the
discriminant capacity of individual parameters (AR, LTFD, F0, clicks) and those
parameters in combination. It is hoped that the findings of the thesis will
encourage discussion leading towards the solution of problems involved in
adopting a numerical LR framework for FSCs. It is also hoped that this research
will prompt forensic phoneticians to consider implementing a verbal LR
framework (or a combination of verbal and numerical LRs) in order to mitigate
the practical limitations associated with completely numerical LRs.
292
Appendix A
Survey Instructions:
Each participant was emailed a unique, secured web-link that directed them to
the survey. The survey had general instructions on the first page, which read:
Please keep in mind that there are no right or wrong answers. This survey is
meant to serve as a tool to gain insight into the general practices in forensic
speaker comparisons around the world as well as finding out which features
forensic phoneticians identify as useful speaker discriminants.
All answers will be kept anonymous and names of participants will never be
revealed, so please answer honestly.
Before each of the 9 sections of the survey there were additional instructions to
remind the participants to generalize to the best of their ability across all cases
they had worked on rather than always responding that the given feature was
case-dependent. The instructions for the 9 sections read as below, with X
representing the number of questions in the section of questions that the
instructions preceded, as detailed in the survey content section of this thesis:
The following section consists of X questions. Please answer the following
questions with regards to all speaker comparison cases you have worked on. I
understand that many features and analyses are case dependent, so please do
your best to make generalizations where applicable.
293
There were some participants who gave the predicted it depends on the case
answer for questions, but for the most part the majority did an excellent job at
generalizing across cases.
294
Appendix B
Example of a Bayesian Network for Speech Evidence:
Bayesian Networks have never been used in FSS before, yet they offer a
method for combining evidence that is both transparent and readily accepted
by other forensic communities (Evett et al., 2002). For this reason, further
research exploring the development of Bayesian Networks for FSC casework
would be worthwhile. A hypothetical example (using the parameters analyzed
in this thesis) is provided below as an illustrative example of a Bayesian
Network using speech evidence, where H represents the hypothesis.
Figure 11.1: Hypothetical Bayesian Network of speech parameters
Articulation
LTFD1 Click Rate F0
Rate
LTFD2 LTFD3 LTFD4
Bayesian Networks, such as the figure above, can be used to calculate Overall
LRs (OLRs) provided that probability densities and variances exist for each
295
parameter. If a case does not account for evidence from a certain parameter in
the figure above, it is possible to simply leave that node out and the remaining
portions of the Network will still be functional for calculating OLRs.
296
List of Abbreviations
change or difference
AR articulation rate
ASR automatic speaker recognition system
AcPA acoustic phonetic analysis
AuPA auditory phonetic analysis
AuPA + AcPA auditory phonetic-cum-acoustic phonetic analysis
Cllr cost log likelihood ratio
CPS classical probability scale
DS different speaker
DyViS Dynamic Variability in Speech
EER equal error rate
F0 fundamental frequency
FSC forensic speaker comparison
FSS forensic speech science
GMM-UBM Gaussian mixture model universal background model
HASR human-assisted automatic speaker recognition
IQR interquartile range
LTFD long-term formant distributions
LTFD1-4 long-term formant distributions one through four
LR likelihood ratio
LLR log10 likelihood ratio
LTS long-term spectrum
297
MFCC Mel-frequency cepstral coefficient
OLR Overall likelihood ratio
SD standard deviation
SS same speaker
SSBE Southern Standard British English
VOT voice onset time
VQ voice quality
298
References
Abercrombie, D. (1967). Elements of General Phonetics. Edinburgh: Edinburgh
University Press.
Aitken, C. G. G. and Gold, E. (2013). Evidence evaluation for discrete

data. Forensic Science International, 230(1-3), pp. 147-155.
Aitken, C. G. G. and Lucy, D. (2004). Evaluation of trace evidence in the form of

multivariate data. Journal of the Royal Statistical Society, 53 (1), pp. 109-122.
Aitken, C. G. G. and Taroni, F. (2004). Statistics and the Evaluation of Evidence for
Forensic Scientists (2nd ed.). Chichester: John Wiley & Sons, Ltd.
Alderman, T. (2004). The Bernard data set as a reference distribution for

Bayesian likelihood ratio-based forensic speaker identification using
formants. Proceedings of the 10th Australian International Conference on
Speech Science & Technology, Macquarie University, Australia., pp. 510-515.
Alwan, A., Narayanan, S. and Haker, K. (1997). Toward articulatory-acoustic

models for liquid approximants based on MRI and EPG data. Part II: the
rhotics. Journal of the Acoustical Society of America, 101(2), pp. 1078-1089.
Atkinson, N. (2009). Formant dynamics for SSBE monophthongs in unscripted

speech. Unpublished; University of York. MSc.
Aristotle. Metaphysica 10f-1045a.
Arndt, H. J. (1963). Untersuchungen ber den Einflu der Mandelausschlung

auf die menschliche Stimme. Folia Phoniatrica, 15, pp. 110-121.
Ash, S. (1988). Speaker identification in sociolinguistics and criminal law. In K.

Ferrara, B. Brown, K. Walters & J. Baugh (eds.) Linguistic Change and Contact.
Austin: University of Texas. pp. 25-33.
Atal, B. S. (1972). Automatic speaker recognition based on pitch contour.

Journal of the Acoustical Society of America, 52(6) (part2), pp. 1687-1697.
Baldwin, J. R. and French, J. P. (1990). Forensic Phonetics. London: Pinter.
Ball, M. J. (1989). Phonetics for speech pathology. London: Whurr Publishers Ltd.
Bauer, H. (1963). Die Beeinflussung der weiblichen Stimme durch androgene

Hormone. Folia Phoniatrica, 15, pp. 264-268.
Bayes, T. and Price, R. (1763). An Introduction to the Doctrine of Fluxions, and

Defence of the Mathematicians against the Objections of the Author of the
299
Analyst. Printed for J. Noon, London. The Eighteenth Century Research
Publications Microfilm A 7173 reel 3774 no. 06.
Becker, T., Jessen, M., and Grigoras, C. (2008). Forensic speaker verification
using formant features and Gaussian mixture models. Proceedings of
Interspeech 2008. Brisbane, pp. 1505-1508.
Berendes, J. (1962). Vernderungen der weiblichen Stimme durch virilisierende

und anabole Hormone. Folia Phoniatrica, 14, pp. 265-271.
Bertsch McGrayne, S. (2012). The Theory that Would Not Die: How Bayes' Rule
Cracked the Enigma Code, Hunted Down Russian Submarines, & Emerged
Triumphant from Two Centuries of Controversy. New Haven and London: Yale
University Press.
Bhuta, T., Patrick, L., and Garnett, J. D. (2004). Perceptual evaluation of voice
quality and its correlation with acoustic measurements. Journal of Voice,
18(3), pp. 299-304.
Boss, D. (1996). F0 and real-life speaker identification. Forensic Linguistics, 3(1),

pp. 155-159.
Bouchayer, M. and Cornut, G. (1992). Microsurgical treatment of benign vocal

fold lesions: Indications, technique, results. Folia Phoniatrica, 44, pp. 155-
184.
Braun, A. (1994). The effect of cigarette smoking on vocal parameters. In

Proceedings of the ESCA Workshop on Automatic Speaker Recognition,
Identification and Verification, Martigny, Switzerland: p. 4.
Braun, A. (1995). Fundamental frequency How speaker-specific is it? In A.

Braun and J.P. Kster (eds.) Studies in Forensic Phonetics. Trier:
Wissenschaftlicher Verlag, pp. 9-23.
Brief Amicus Curiae of Americans for Effective Law Enforcement (1997).

Submitted in Kumho Tire v. Carmichael, No. 97-1709, in the Supreme Court
of the United States (October Term).
Broeders, A. P. A. (1999). Some observations on the use of probability scales in

forensic identification. International Journal of Speech, Language and the Law,
6(2), pp. 228-241.
Broeders, A. P. A. (2001). Forensic Speech and Audio Analysis. Forensic

Linguistics. 1998 to 2001 A Review, Proc. 13th INTERPOL Forensic Sciences
Symposium. 16-19 October. Lyon, France.
Brmmer, N., Burget, L., Cernock, J. H., Glembek, O., Grzl, F., Karafit, M., van
Leeuwen, D. A., Matejka, P., Schwarz, P., and Strasheim, A. (2007). Fusion of
heterogeneous speaker recognition systems in the STBU submission for the
300
NIST SRE 2006. IEEE Transactions on Audio Speech and Language Processing,
15, pp. 2072-2084.
Brmmer, N. and du Preez, J. (2006). Application-independent evaluation of

speaker detection. Computer Speech and Language, 20(2-3), pp. 230-275.
Byrne, C. and Foulkes, P. (2004). The mobile phone effect on vowel formants.
International Journal of Speech, Language and the Law, 11, pp. 83-102.
Cambier-Langeveld, T. (2007). Current methods in forensic speaker

identification. International Journal of Speech, Language and the Law, 14(2),
pp. 224-243.
Campbell, J. (1997). Speaker recognition: a tutorial. Proceedings of the IEEE, 85

(9), pp. 1437-1462.
Campbell, J., Shen, W., Campbell, W., Schwartz, R., Bonastre, J. F., and Matrouf, D.
(2009). Forensic speaker recognition: A need for caution. IEEE Signal
Processing Magazine, March 2009, pp. 95-103.
Cao, H. and Wang, Y. (2011). A forensic aspect of articulation rate variation in

Chinese. Proceedings of the 17th International Congress of Phonetic Sciences,
August 17-21, 2011, Hong Kong, China, pp. 396-399.
Chambers, J. K. (2005). Sociolinguistic theory (2nd ed). Oxford, UK: Blackwell.
Champod, C. and Evett, I. W. (2000). Commentary on Broeders 1999. Forensic

Linguistics, 7(2), pp. 238-243.
Champod, C., & Meuwly, D. (2000). The inference of identity in forensic speaker
recognition. Speech communication, 31, pp. 193-203.
Clark, J., & Yallop, C. (2001). An Introduction to Phonetics and Phonology (2nd
ed.). Oxford: Blackwell Publishers Ltd.
Clermont, F., French, P., Harrison, P., and Simpson, S. (2008). Population data for
English spoken in England: A modest first step. Paper presented at the
Conference of the International Association of Forensic Phonetics and
Acoustics. Plymouth, United Kingdom.
Clermont, F., Harrison, P., and French, P. (2007). Formant-pattern estimation

guided by cepstral compatibility. Paper presented at the Conference of the
International Association of Forensic Phonetics and Acoustics. Plymouth,
United Kingdom.
Coe, P. (2012). The effect of sample contemporaneity on the outcome of likelihood

ratios for forensic voice comparison. Unpublished; University of York. MSc.
301
Crystal, D. (1987). The Cambridge Encyclopaedia of Language. Cambridge:
Cambridge University Press.
Damast, P. H. (1964). Virilization of the voice due to anabolic steroids. Folia

Phoniatrica, 16, pp. 10-18.
Damast, P. H. (1967). Voice change in adult women caused by virilizing agents.

Journal of Speech and Hearing Disorders, 32, pp. 126-132.
Dankoviov J. (1997). The domain of articulation rate variation in Czech.

Journal of Phonetics, 25, pp. 287-312.
Darby, J. K. and Hollien, H. (1977). Vocal and speech patterns of depressive

patients. Folia Phoniatrica, 29, pp. 279-291.
DeGroot, M. and Fienberg S. (1983). The Comparison and Evaluation of

Forecasters. Journal of the Royal Statistical Society. Series D (The Statistician),
Vol. 32 (1), pp. 12-22.
de Jong, G., McDougall, K., and Nolan, F. (2007) Sound Change and Speaker
Identity: An Acoustic Study. In: Christian Mller (ed.), Speaker Classification
II: Selected Papers. Berlin: Springer, pp. 130-141.
Dieroff, H. G. and Siegert, C. (1966). Tonhhenverschiebung unter

Lrmbelastung. Folia Phoniatrica, 24, pp. 247-255.
Doherty, R. and Lee, A. (2009). Speech rates of Irish English-speaking adults.

The Royal College of Speech and Language Therapists Scientific Conference
(RCSLT), London, UK.
Drygajlo, A. (2007). Forensic automatic speaker recognition. IEEE Signal

Processing Magazine, 24, pp. 132-135
Eckert, P. (2000). Linguistic Variation as Social Practice. Oxford, UK: Blackwell.
Enzinger, E. (2010a). Characterizing formant tracks in Viennese diphthongs for

forensic speaker comparison. Proceedings of the AES 39th International
Conference Audio Forensics, Hillerd, Denmark, pp. 4752.
Enzinger, E. (2010b). Measuring the Effects of the Adaptive Multi-Rate (AMR)

codecs on formant tracker performance. Paper presented at the conference of
the Second Pan-American/Iberian Meeting of the Acoustics for the Acoustical
Society of America, Cancn, Mxico.
Enzinger, E. and Morrison, G. S. (2012). The importance of using between-

session test data in evaluating the performance of forensic-voice-comparison
systems. In Proceedings of the 14th Australasian International Conference on
Speech Science and Technology, pp. 137-140.
302
Evett, I. W. (1990). The theory of interpreting scientific transfer evidence.
Forensic Science Progress, 4, pp. 141179.
Evett, I. W. (1995). Avoiding the transposed conditional. Science and Justice, 35

(2), pp. 127-131.
Evett, I. W., Gill, P. D., Jackson, G., Whitaker, J., and Champod, C. (2002).
Interpreting small quantities of DNA: The hierarchy of propositions and the
use of Bayesian networks. Journal of Forensic Sciences, 47(3), pp. 520-530.
Evett, I. W., Jackson, G. Lambert, J. A., and McCrossan, S. (2000). The impact of
the principles of evidence interpretation on the structure and content of
statements. Science & Justice, 40, pp. 233239.
Evett, I. and Weir, B. S. (1998). Interpreting DNA Evidence: Statistical Genetics for
Forensic Scientists. Sunderland, Massachusetts: Sinauer Associates.
Foulkes, P. and French, P. (2012). Forensic phonetic speaker comparison. In

Solan, L. and Tiersma, P. (eds.) Oxford Handbook of Language and Law.
Oxford: Oxford University Press, pp. 557-572.
French, P. (1990). Acoustic Phonetics. In Baldwin, J. and French, P. (eds.)

Forensic Phonetics. London: Pinter Publishers. pp. 42-64.
French, J. P. (1994). An overview of forensic phonetics with particular reference

to speaker identification. Forensic Linguistics, 5(1), pp. 58-68.
French, P., Foulkes, P., Harrison, P., and Stevens, L. (2012). Vocal tract output
measures: relative efficacy, interrelationships and limitations. Paper
presented at the Conference of the International Association of Forensic
Phonetics and Acoustics Conference, Santander, Spain.
French, J. P. and Harrison, P. (2007). Position statement concerning use of

impressionistic likelihood terms in forensic speaker comparison cases.
International Journal of Speech, Language and the Law, 14(1), pp. 137-144.
French, P. and Harrison, P. (2010). The work of speech and audio analysts. AES
Conference no. 128. London, United Kingdom.
French, J. P., Nolan, F., Foulkes, P., Harrison, P., and McDougall, K. (2010). The
UK position statement on forensic speaker comparison: a rejoinder to Rose
and Morrison. International Journal of Speech, Language and the Law, 17(1),
pp. 143-152.
French, P. and Stevens, L. (2013). Forensic speech science. In Jones, M. & R.

Knight (eds.), The Bloomsbury Companion to Phonetics. London: Bloomsbury,
pp. 183-197.
303
Fritzell, B., Sundberg, J., and Strange-Ebbesen, A. (1982). Pitch change after
stripping oedematous vocal folds. Folia Phoniatrica, 34, pp. 29-32.
Garrett, K. L. and Healey, E. Ch. (1987). An acoustic analysis of fluctuations in

the voices of normal adult speakers across three times of day. Journal of the
Acoustical Society of America, 82, pp. 58-62.
Gilbert, H. R. and Weismer, G. G. (1974). The effects of smoking on the speaking

fundamental frequency of adult women. Journal of Psycholinguistic Research,
3, pp. 225-231.
Giles, H. (1973). Accent mobility: A model and some data. Anthropological

Linguistics, 15(2), pp. 87105.
Giles, H. and Ogay, T. (2007). Communication accommodation theory. In

Whaley, B. B. & W. Samter (eds.), Explaining communication: Contemporary
theories and exemplars. Mahwah, NJ: Lawrence Erlbaum Associates, pp. 293-
310.
Gimson, A. C. (1970). An Introduction to the Pronunciation of English. Second

Edition. London: Edward Arnold.
Goel, V. (2000). Minimum Bayes-risk automatic speech recognition. Computer

Speech and Language, 14 (2), pp. 115-135.
Gold, E. (2009). The Effects of Video and Voice Recorders in Cellular Phones on
Vowel Formants and Fundamental Frequency. Unpublished; University of
York. MSc.
Gold, E. and Hughes, V. (2012). Defining interdependencies between speech

parameters. BBfor2 Short Summer School in Forensic Evidence Evaluation and
Validation. Madrid, Spain.
Gold, E. and Hughes, V. (2013). Identifying correlations between speech

parameters for forensic speaker comparisons. Research grant provided by the
International Association of Forensic Phonetics and Acoustics.
[http://www.iafpa.net/GoldHughes_CorrelationsBetweenSpeech
Parameters.html]
Goldman-Eisler, F. (1956). The Determinants of the Rate of Speech Output and

their Mutual Relations. Journal of Psychosomatic Research, Vol. 1, pp 137-143.
Goldman-Eisler, F. (1968). Psycholinguistics: Experiments in Spontaneous Speech.

Academic Press: London.
Gonzalez-Rodriguez, J., Rose, P., Ramos D., Toledano, D. T., and Ortega-Garcia, J.
(2007). Emulating DNA: rigorous quantification of evidential weight in
304
transparent and testable forensic speaker recognition, IEEE Transactions of
Audio, Speech and Language Processing, 15, pp. 2104-2115.
Graddol, D. (1986). Discourse specific pitch behavior. In Johns-Lewis, C. (ed.).

Intonation in Discourse. London: Croom Helm, pp. 221-237.
Greenberg, J. H. (1950). Studies in African linguistic classification VI: the click

languages. Southwestern Journal of Anthropology, 6 (3), pp. 223-237.
Greenberg, C. S., Martin, A., Brandschain, L., Campbell, J., Cieri, C., Doddington, G.,
and Godfrey, J. (2010). Human Assisted Speaker Recognition In SRE10, Proc.
Odyssey 2010, Brno, Czech Republic, June-July 2010.
Grigoras, C. (2001). Digital Voice Processing System. Unpublished; University of

Bucharest. PhD.
Grigoras, C. (2003). Voice analysis on noisy recordings. Paper presented at the

Cambridge Forensic Phonetics Workshop, August, Cambridge, UK.
Guillemin, B. and Watson, C. (2008). Impact of the GSM mobile phone network
on the speech signal: some preliminary findings. International Journal of
Speech, Language and the Law, 15(2), pp. 193-218.
Halliday, M. A. K. (1967). Intonation and Grammar in British English. The Hague:

Mouton.
Hand, D. J., and Yu, K. (2001). Idiot's Bayes - not so stupid after all? International
Statistical Review, 69(3), pp. 385-399.
Hardcastle, W. J. (1975). Some aspects of speech production under controlled

conditions of oral anaesthesia and auditory masking. Journal of Phonetics, 3,
pp. 197-214.
Harrison, P. (2004). Variability of Formant Measurements. Unpublished;

University of York. MA.
Hecker, M. L., Stevens, K. N., von Bismarck, G., and Williams, C. E. (1968).
Manifestations of task-induced stress in the acoustic speech signal. Journal of
the Acoustical Society of America, 44, pp. 993-1001.
Henze, R. (1953). Experimentelle Untersuchungen zur Phnomenologie der

Sprechgeschwindigkeit. Zeitschrift fr Experimentelle und Angewandte
Psychologie, 1, pp. 214-243.
Herbert, R.K. (1990) The sociohistory of clicks in Southern Bantu.

Anthropological Linguistics, 32 (3), pp. 295-315.
Hodgson D. (2002) A lawyer looks at Bayes theorem. The Australian Law

Journal, 76, pp. 109 118.
305
Hollien, H. (1980). Vocal indicators of psychological stress. In Wright, F., Bahn,
C., and Rieben, R.W. (eds.) Forensic Psychology and Psychiatry. New York:
John Wiley & Sons Ltd., pp. 47-72.
Hollien, H. and Majewski, W. (2009). Unintended Consequences: Due to Lack of

Standards for Speaker Identification and Other Forensic Procedures. Paper
presented at the Sixteenth International Congress on Sound and Vibration,
Krakow, Poland, pp. 1-6
Hollien, H. and Michel, J. F. (1968). Vocal fry as a phonational register. Journal of

Speech Language, and Hearing Research, 11, pp. 600-604.
Horii, Y. (1975). Some statistical characteristics of voice fundamental frequency

and chronological age in males. Journal of Speech, Language, and Hearing
Research, 18, pp. 192-201.
House of Commons Northern Ireland Affairs Committee (2009) Cross-border Co-

operation between the Governments of the United Kingdom and the Republic of
Ireland: Second Report of Session 20082009. London: The Stationery Office.
Hudson, T., de Jong, G., McDougall, K., Harrison, P., and Nolan, F. (2007). F0
statistics for 100 young male speakers of Standard Southern British English.
In 16th Proceedings of the International Congress of Phonetic Sciences,
Saarbrcken, pp. 1809-1812.
Hudson, A. I. and Holbrook, A. (1981). A study of the reading fundamental vocal

frequency of young black adults. Journal of Speech, Language, and Hearing
Research, 24, pp. 197-201.
Hughes, V. (in progress). The Effects of Variability on the Outcome of Numerical

Likelihood Ratios. Unpublished; University of York. PhD.
Hughes, V. (2011). The Effect of Variability on the Outcome of Likelihood Ratios.

Unpublished; University of York. MSc.
Hughes, V. (2013). Establishing typicality: a closer look at individual formants.

In Proceedings of Meetings on Acoustics, POMA Volume 19, pp. 060042-
060050.
Hughes, V. and Foulkes, P. (2012). Effects of variation on the computation of

numerical likelihood ratios for forensic voice comparison. Paper presented at
the Conference of the International Association of Forensic Phonetics and
Acoustics. Universidad Internacional Menndez Pelayo, Santander.
Hughes, V., McDougall, K. and Foulkes, P. (2009). Diphthong dynamics in

unscripted speech. Paper presented at the Conference of the International
Association of Forensic Phonetics and Acoustics. Cambridge, UK. 2-5 August
2009.
306
Ishihara, S. and Kinoshita, Y. (2008). How many do we need? Exploration of the
Population Size Effect on the performance of forensic speaker classification.
9th Annual Conference of the International Speech Communication Association
(Interspeech). Brisbane, Australia, pp. 1941-1944.
Jacewicz E. Fox R. A. ONeil C. and Salmons J. (2009) Articulation rate across

dialect, age and gender. Language Variation and Change, 21, pp. 233-256.
Jessen, M. (1997). Speaker-specific information in voice quality parameters.

Forensic Linguistics, 4(1), pp. 84-103.
Jessen, M. (2007a). Speaker Classification in Forensic Phonetics and Acoustics.

In Mller, C. (ed.), Speaker classification I: Fundamentals, features, and
methods, Berlin, Germany: Springer, pp. 180-204.
Jessen, M. (2007b). Forensic reference data on articulation rate in German.

Science and Justice, 47, pp. 5067.
Jessen, M. (2008). Forensic phonetics. Language and Linguistics Compass, 2(4),

pp. 671-711.
Jessen, M. and Becker, T. (2010). Long-term formant distribution as a forensic-

phonetic feature. Paper presented at the Conference of the Acoustical Society
of America. Cancun, Mexico.
Jessen, M. and Roux, J.C. (2002). Voice quality differences associated with stops
and clicks in Xhosa. Journal of Phonetics, 30, pp. 152.
Jessen, M., Enzinger, E., and Jessen, M. (2013). Experiments on Long-Term

Formant Analysis with Gaussian Mixture Modeling using VOCALISE. Paper
presented at the Conference of the International Association of Forensic
Phonetics and Acoustics, Tampa, Florida, USA.
Jessen, M., Kster, O., and Gfroerer, S. (2005). Influence of vocal effort on
average and variability of fundamental frequency. International Journal of
Speech, Language, and the Law, 12(2), pp. 174-213.
Johnson, K. (2003). Acoustic and auditory phonetics. (2nd edn). Malden, MA:
Blackwell.
Junqua, J. (1996) The influence of acoustics on speech production: A noise-

induced stress phenomenon known as the Lombard reflex. Speech
Communication, 20, pp. 13-22.
Kavanagh, C. (2010). Speaker discrimination using English nasal durations and

formant dynamics. Paper presented at the Conference of the International
Association of Forensic Phonetics and Acoustics. Trier, Germany.
307
Kavanagh, C. (2011). Intra- and inter-speaker variability in duration and
spectral properties of English /s/. Paper presented at the Conference of the
Acoustical Society of America, San Diego, USA.
Kavanagh, C. (2013). New Consonantal Acoustic Parameters for Forensic Speaker

Comparison. Unpublished; University of York. PhD.
Kaye, D. H. (1993). DNA evidence: Probability, population genetics, and the

courts. Harvard Journal of Law and Technology, 7(1), pp. 101-172.
Keating, P. (1988). A Survey of Phonological Features. Indiana University

Linguistics Club, Bloomington.
Keilmann, A. and Hlse, M. (1992). Dysphonie nach Strumektomie bei

ungestrter respiratorischer Beweglichkeit der Stimmlippen. Folia
Phoniatrica, 44, pp. 261-268.
Kinoshita, Y. (2001). Testing Realistic Forensic Speaker Identification in Japanese:

A Likelihood Ratio-Based Approach Using Formants. Unpublished; Australian
National University. PhD.
Kinoshita, Y. (2002). Use of likelihood ratio and Bayesian approach in forensic

speaker identification. Proceedings of the 9th Australian International
conference on Speech Science and Technology. Melbourne, Australia, pp. 297-
302.
Kinoshita Y. (2005). Does Lindleys LR estimation formula work for speech

data? Investigation using long-term F0. International Journal of Speech,
Language and the Law, 12, pp. 235-254.
Kinoshita, Y. and Ishihara, S. (2012). The effect of sample size on the

performance of likelihood ratio-based forensic voice comparison.
Proceedings of the 14th Australasian International Conference on Speech
Science and Technology. Macquarie University, Australia.
Kinoshita, Y., Ishihara, S., and Rose, P. (2009). Exploring the discriminatory
potential of F0 distribution parameters in traditional forensic speaker
recognition. International Journal of Speech, Language and the Law, 16, pp.
91-111.
Kinoshita, Y. and Osanai, T. (2006). Within-speaker variation in diphthongal

dynamics: what can we compare? Proceedings of the 11th Australasian
International Conference on Speech Science and Technology, University of
Auckland, New Zealand, pp. 112-117.
Kirchhbel, C. (2010). The effects of Lombard speech on vowel formant

measurements. Paper presented at the Conference of the Acoustical Society of
America, Cancun, Mexico.
308
Kirchhbel, C. and Howard, D. (2011). Investigating the acoustic characteristics
of deceptive speech. Proceedings of the 17th International Congress of
Phonetic Sciences, August 17-21, 2011, Hong Kong, China, pp. 1094-1097.
Klingholz, F., Penning, R., and Liebhardt, E. (1988). Recognition of low-level

alcohol intoxication from speech signal. Journal of the Acoustical Society of
America, 84, pp. 929-935.
Kononenko, I. (1990). Comparison of inductive and nave Bayesian capitalised

learning approaches to automatic knowledge acquisition. In B. Wielinga et al.
(Eds.) Current trends in knowledge acquisition. IOS Press: Amsterdam,
Netherlands, pp. 190-197.
Konrat, C. and Jessen, M. (2013). Fundamental Frequency Analysis: A

Collaborative Exercise. Paper presented at the Conference of the International
Association of Forensic Phonetics and Acoustics, Tampa, Florida, USA.
Koreman, J. (2006). Perceived speech rate: the effects of articulation rate and
speaking rate in spontaneous speech. Journal of the Acoustical Society of
America, 119, pp. 582-596.
Kuhn, T. S. (1962). The Structure of Scientific Revolutions. Chicago: University of

Chicago Press.
Knzel, H. J. (1987). Sprechererkennung: Grundzge forensischer

Sprachverarbeitung. Heidelberg: Kriminalistik Verlag.
Knzel, H. J. (1989). How well does average fundamental frequency correlate

with speaker height and weight? Phonetica, 46, p. 117.
Knzel, H. (1997). Some general phonetic and forensic aspects of speaking

tempo. Forensic Linguistics, 4, pp. 48-83.
Knzel, H. J. (2000). Effects of voice disguise on speaking fundamental

frequency. Forensic Linguistics, 7, pp. 149-179.
Knzel H. (2001). Beware of the telephone effect the influence of telephone

transmission on the measurement of formant frequencies. International
Journal of Speech, Language and the Law, 8, pp. 80-99.
Knzel, H., Braun, A., and Eysholdt, U. (1992). Einflu von Alkohol auf Stimme
und Sprache. Heidelberg: Kriminalistik-Verlag.
Ladd, D. R. (1996). Intonational phonology. Cambridge: Cambridge University

Press.
Ladefoged, P. (2006). A Course in Phonetics, 5th edn. Boston: Wadsworth

Cengage Learning.
309
Laplace, P. S. (1781). Mmoire sur les Probabilitis. OC (9), pp. 383-485.
Laver, J. (1980). The Phonetic Description of Voice Quality. Cambridge:

Laver, J. (1994). Principles of Phonetics. Cambridge: Cambridge University Press.
Law Commission of England and Wales (2011). Expert Evidence in Criminal

Proceedings in England and Wales. No. 325
Linard, J. and Di Benedetto, M. (199). Effect of vocal effort on spectral

properties of vowels. Journal of the Acoustical Society of America, 106(1), pp.
411-422.
Lindh J. (2006) Preliminary Descriptive F0-statistics for Young Male Speakers

Lund University Working Papers, 52, pp. 89-92.
Lindh, J. (2007). Fundamental frequency and the alternative baseline in forensic

speaker identification. Paper presented at the Conference of the International
Association for Forensic Phonetics and Acoustics (IAFPA), Plymouth, UK.
Lindh, J., & Morrison, G.S. (2011). Humans versus machine: Forensic voice
comparison on a small database of Swedish voice recordings. Proceedings of
the 17th International Congress of Phonetic Sciences, Hong Kong, China, pp.
12541257.
Lindley, D. V. (1977). A problem in forensic science. Biometrika, 64, pp. 207-213.
Loakes, D. (2006). Variation in long-term fundamental frequency:

measurements from vocalic segments in twins speech. Proceedings of the
11th Speech Science and Technology Conference, Auckland, New Zealand, pp.
205-210.
Lombard E. (1911). Le signe de llvation de la voix. Annales des maladies de

loreille et du larynx, 37, pp. 101-119.
Lucy, D. (2005). Introduction to Statistics for Forensic Scientists. Chichester: John

Wiley & Sons, Ltd.
Mead, K. O. (1974). Identification of speakers from fundamental-frequency

contours in conversational speech. Joint Speech Research Unit, Report No.
1002. Ruislip, Middlesex.
Miller, J. L., Grosjean, F., and Lomanti, C. (1984). Articulation rate and its
variability in spontaneous speech: a reanalysis and some implications.
Phonetica, 41, pp. 215-225.
310
McDougall, K. (2004). Speaker-specific formant dynamics: an experiment on
Australian English /a/. International Journal of Speech, Language and the
Law, 11 (1), pp. 103-130.
McDougall, K. (2006). The role of formant dynamics in determining speaker

identity. Ph.D. abstract. International Journal of Speech, Language and the
Law, 13(1), pp. 144-145.
McDougall, K. and Nolan, F. (2007). Discrimination of speakers using formant

dynamics of /u:/ in British English. Proceedings of the 16th International
Congress of Phonetic Sciences. Saarbrcken, Germany, pp. 1445-1448.
Milroy, L. and Gordon, M. (2003). Sociolinguistics: Method and Interpretation.

Oxford: Blackwell.
Moos, A. (2010). Long-term formant distribution as a measure of speaker

characteristics in read and spontaneous speech. The Phonetician, 101, pp. 7-
24.
Morrison, G. S. (2007). MatLab implementation of Aitken and Lucys (2004)

forensic likelihood ratio software using multivariate-kernel-density
estimation. Downloaded: December 2011.
Morrison, G. S. (2008) Forensic voice comparison using likelihood ratios based

on polynomial curves fitted to the formant trajectories of Australian English
/a/. International Journal of Speech, Language, and the Law, 15(2), pp. 249-
266.
Morrison, G. S. (2009a). Likelihood-ratio voice comparison using parametric

representations of the formant trajectories of diphthongs. Journal of the
Acoustical Society of America, 125, pp. 2387-2397.
Morrison, G. S. (2009b). The place of forensic voice comparison in the ongoing

paradigm shift. Written version of an invited presentation given at the 2nd
International Conference on Evidence Law and Forensic Science. 25-26 July,
Beijing, China, pp. 1-16.
Morrison, G. S. (2009c). Forensic voice comparison and the paradigm shift.

Science and Justice, 49, pp. 298-308.
Morrison, G. S. (2011). A comparison of procedures for the calculation of

forensic likelihood ratios from acousticphonetic data: Multivariate kernel
density (MVKD) versus Gaussian mixture modeluniversal background
model (GMMUBM). Speech Communication, 53, pp. 242-256.
Morrison, G. S. (2012) Tutorial on logistic-regression calibration and fusion:

Converting a score to a likelihood ratio. Australian Journal of Forensic
Sciences, 45 (2), pp.173-197.
311
Morrison, G. S. (2013). Tutorial on logistic-regression calibration and fusion:
converting a score to a likelihood ratio. Australian Journal of Forensic
Sciences, 45, pp. 173-197.
Morrison, G. S. and Kinoshita, Y. (2008). Automatic-type calibration of

traditionally derived likelihood ratios: Forensic analysis of Australian English
/o/ formant trajectories. Proceedings of Interspeech 2008 International
Speech Communication Association, pp. 15011504.
Morrison, G. S., Ochoa, F., and Thiruvaran, T. (2012). Database selection for
forensic voice comparison. Proceedings of Odyssey 2012: The Language and
Speaker Recognition Workshop, Singapore, International Speech
Communication Association, pp. 62-77.
Morrison, G. S., Rose, P., & Zhang, C. (2012). Protocol for the collection of
databases of recordings for forensic-voice-comparison research and practice.
Australian Journal of Forensic Sciences, 44, pp. 155167.
Morrison, G.S., Thiruvaran, T., and Epps, J. (2010). An issue in the calculation of
logistic-regression calibration and fusion weights for forensic voice
comparison. Proceedings of the 13th Australasian International Conference on
Speech Science and Technology. Melbourne, Australia, pp. 74-77.
Murphy, C. H. and Doyle, P. C. (1987). The effects of cigarette smoking on voice-

fundamental frequency. Otolaryngology-Head and Neck Surgery, 97, pp. 376-
380.
Narang, V., Misra, D. and Yadav, R. (2012). F1 and F2 correlation with F0: A
study of vowels of Hindi, Punjabi, Korean, and Thai. International Journal of
Asian Language Processing, 22(2), pp. 63-73.
National Research Council (2009). Strengthening Forensic Science in the United

States: A Path Forward. Washington D.C.: The National Academic Press.
Nolan, F. (1983). The Phonetic Bases of Speaker Recognition. Cambridge:

Nolan, F. and Grigoras, C. (2005). A case for formant analysis in forensic speaker
identification. International Journal of Speech, Language and the Law, 12(2),
pp. 143-173.
Nolan, F., McDougall, K., de Jong, G. and Hudson, T. (2009). The DyViS database:
style-controlled recordings of 100 homogeneous speakers for forensic
phonetic research. International Journal of Speech, Language and the Law,
16(1), 31-57.
Novak, A., Dlouha, O., Capkova, B., and Vohradnik, M. (1991). Voice fatigue after
theater performance in actors. Folia Phoniatrica, 43, pp. 74-78.
312
Oats, J. M. and Dacakis, G. (1983). Speech pathology considerations in the
management of transsexualism a review. British Journal of Disorders of
Communication, 18, pp. 139-151.
Ogden, R. (2013). Clicks and percussives in English conversation. Journal of the

International Phonetic Association, 43(3), pp. 299-320.
Osterburg, J. W. (1969). The evaluation of physical evidence in criminalistics:

Subjective or objective process?, Journal of Criminal Law and Criminology, 60,
pp. 97-101.
Papp, V. (2008). The Effects of Heroin on Speech. Unpublished; University of

York. MSc.
Peterson, G. E. (1959). The acoustics of speech part II: acoustical properties of

speech waves. In Travis, L. E. (ed.) Handbook of Speech Pathology. London:
Peter Owen, pp. 137-173.
Pike, K. (1943). Phonetics: A critical analysis of phonetic theory and a technic for
the practical description of sounds. Ann Arbor, MI: University of Michigan
Press.
Pisoni, D. B. and Martin, C. S. (1989). Effects of alcohol on the acoustic-phonetic

properties of speech: perceptual and acoustic analysis. Alcoholism: Clinical
and Experimental Research, 13, pp. 577-587.
Porwal, U., Shi, Z., and Setlur, S. (2013). Machine learning in handwritten Arabic
text recognition. In Govindaraju, V. and C. R. Rao (eds.) Handbook of Statistics
31 - Machine Learning: Theory and Applications. Amsterdam: North Holland,
pp. 443-470.
Ramos-Castro, D. (2007). Forensic evaluation of the evidence using automatic

speaker recognition systems. Unpublished PhD Dissertation, Universidad
Autnoma de Madrid.
Ramos-Castro, D., Gonzalez-Rodriguez, J. and Ortega-Garcia, J. (2006).

Likelihood Ratio Calibration in a Transparent and Testable Forensic Speaker
Recognition Framework. In the Proceedings of the IEEE Speaker and
Language Recognition Workshop, pp. 1-8.
Reynolds, D. A. (2002). An overview of automatic speaker recognition

technology. Acoustics, Speech, and Signal Processing (ICASSP), Proceedings of
the IEEE International Conference 4, pp. 4072-4075.
Reynolds, D. A., Quatieri, T. F., and Dunn, R. (2000). Speaker verification using
adapted Gaussian mixture models. Digital Signal Processing, 10 (1-3), pp. 19-
41.
313
Rhodes, R. (2013). Assessing the Strength of Non-Contemporaneous Forensic
Speech Evidence. Unpublished; University of York. PhD.
Robb, M., Maclagan, M. A., and Chen, Y. (2004) Speaking rates of American and
New Zealand varieties of English. Clinical Linguistics & Phonetics, 18(1), pp. 1-
15.
Robertson, B. and Vignaux, G. A. (1995). Interpreting evidence: evaluating

forensic science in the courtroom. Chichester: John Wiley & Sons, Ltd.
Rose, P. (1999). Long- and short-term within-speaker differences in the

formants of Australian hello. Journal of the International Phonetic Association,
29(1), pp. 1-31.
Rose, P. (2002). Forensic speaker identification. London: Taylor-Francis Ltd.
Rose, P. (2003). The technical comparison of forensic voice samples. In I.S.

Freckleton and H. Selby (eds.) Expert Evidence. North Ryde: Lawbook Co, Ch.
99.
Rose, P. (2006a). The intrinsic forensic discriminatory power of diphthongs.

Proceedings of the 11th Australasian International Conference on Speech
Science and Technology. University of Auckland, New Zealand, pp. 64-69.
Rose, P. (2006b). Technical forensic speaker recognition: evaluation, types and

testing of evidence. Computer Speech and Language, 20, pp. 15991.
Rose, P. (2006c). Accounting for correlation in linguistic-acoustic likelihood

ratio-based forensic speaker discrimination. Proceedings of the Speaker and
Language Recognition Workshop, pp. 1- 8.
Rose, P. (2007a). Forensic speaker discrimination with Australian English vowel

acoustics. Proceedings of the 16th International Congress of Phonetic Sciences.
Saarbrcken, Germany, pp. 1817-1820.
Rose, P. (2007b). Going and getting it Forensic Speaker Recognition from the
perspective of a traditional practitioner/researcher. Paper presented at the
Australian Research Council Network in Human Communicative Science
Workshop: FSI not CSI Perspectives in State-of-the-Art Forensic Speaker
Recognition, Sydney.
Rose P. (2010a). Bernards 18 vowel inventory size and strength of forensic

voice comparison evidence. Proceedings of the 13th Australian International
Conference on Speech and Technology. Melbourne, Australia, pp. 30-33.
Rose, P. (2010b). The effect of correlation on strength of evidence estimates in

forensic voice comparison: uni- and multivariate likelihood ratio-based
discrimination with Australian English vowel acoustics. International Journal
of Biometrics, 2, pp. 316-329.
314
Rose, P. (2011). Forensic voice comparison with Japanese vowel acoustics a
likelihood ratio-based approach using segmental cepstra. In W. S. Lee, E. Zee
(Eds.) Proceedings of the 17th International Congress of Phonetic Sciences,
Hong Kong, 17-21 August, pp. 1718-1721.
Rose, P. (2012). Yes, Not Too Bad Likelihood Ratio-Based Forensic Voice
Comparison in a $150 Million Telephone Fraud. Proceedings of the 14th
Australasian International Conference on Speech Science and Technology.
Macquarie University, Australia, pp. 161-164.
Rose, P. (2013a). More is better: likelihood ratio-based forensic voice

comparison with vocalic segmental cepstra frontends. International Journal
of Speech, Language and the Law, 20(1), pp. 77-116.
Rose, P. (2013b). Where the science ends and the law begins: Likelihood ratio-
based forensic voice comparison in a $150 million telephone fraud.
International Journal of Speech, Language and the Law, 20(2), pp. 277-324.
Rose, P., Kinoshita, Y., and Alderman, T. (2006). Realistic extrinsic forensic
speaker discrimination with the diphthong /a/. Proceedings of the 11th
Australasian International Conference on Speech Science and Technology.
University of Auckland, New Zealand, pp. 329-334.
Rose, P., Lucy, D., and Osanai, T. (2004). Linguistic-acoustic forensic speaker
identification with likelihood ratios from a multivariate hierarchical random
effects model a non-idiots Bayes approach. Proceedings of the 10th
Australasian Conference on Speech Science and Technology. Macquarie
University, Australia, pp. 492-497.
Rose, P. and Morrison, G. S. (2009). A response to the UK position statement on

forensic speaker comparison. International Journal of Speech, Language and
the Law, 16 (1), pp. 139-163.
Rose, P., Osanai, T., and Kinoshita, Y. (2003). Strength of forensic speaker
identification evidence: multispeaker formant- and cepstrum-based
segmental discrimination with a Bayesian likelihood ratio as threshold.
Forensic Linguistics, 10, pp. 179-202.
Rose, P. and Winter, E. (2010). Traditional forensic voice comparison with

female formants: Gaussian mixture model and multivariate likelihood ratio
approaches. Proceedings of the 13th Australasian International Conference on
Speech Science and Technology. Melbourne, Australia, pp. 42-45.
Saks, M. J. and Koehler, J. J. (2005). The coming paradigm shift in forensic

identification science. Science, 309, pp. 892-895.
315
Saxman, J. H. and Burk, K. W. (1968). Speaking fundamental frequency
characteristics of adult female schizophrenics. Journal of Speech, Language,
and Hearing Research, 11, pp. 194-203.
Scherer, K. R. (1977). The effect of stress on fundamental frequency of the voice.

Journal of the Acoustical Society of America, 6, pp. 25-26.
Scherer, K. R., Helfrich, H., Standke, R., and Wallbott, H. (1976). Psychoakustische
und kinesische Verhaltensanalyse. Research report, University of Giessen.
Schultz-Coulon, H. J. (1975). Bestimmung und Beurteilung der individuellen

mittleren Sprechstimmlage. Folia Phoniatrica, 27, pp. 375-386.
Schultz-Coulon, H. J. and Fues, C. P. (1976). Der Lombard-Reflex als

Stimmfunktionsprfung. HNO 24, pp. 200-204.
Shepard, C. A., Giles, H. and LePoire, B. (2001). Communication Accommodation

Theory. In, Robinson, W. Peter & Howard Giles (eds.), The New Handbook of
Language and Social Psychology. Chichester: Wiley, pp. 34-51.
Simpson, S. (2008). Testing the Speaker Discrimination Ability of Formant

Measurements in Forensic Speaker Comparison Cases. Unpublished; University
of York. MSc.
Sobell, L. C., Sobell, M. B., and Coleman, R. F. (1982). Alcohol-induced dysfluency

in nonalcoholics. Folia Phoniatrica, 34, pp. 316-323.
Sorensen, D. and Horii, Y. (1982). Cigarette smoking and voice fundamental

frequency. Journal of Communication Disorders, 15, pp. 135-144.
Steffan-Batog, M., Jassem, W., and Gruszka-Koscielak, H. (1970). Statistical

distributions of short term F0 values as a personal voice characteristic. In
Jassem, W. (ed.) Speech Analysis and Synthesis. Warsaw: Polish Academy of
Sciences, Volume 2: 197-208.
Stevens, K. N. (2001). Acoustic Phonetics. Cambridge, MA: MIT Press.
Stevens, L. and French, P. (2012). Voice quality in standard southern British

English: distribution of features, inter-speaker variability, and the effect of
telephone transmission. Paper presented at the Conference of the
International Association of Forensic Phonetics and Acoustics. Santander,
Spain.
Stoney, D. A. (1991). What made us ever think we could individualize using

statistics? Journal of the Forensic Science Society, 31(2), pp. 197
SurveyGizmo (2006). International Survey on Forensic Speaker Comparison

Practices. http://www.surveygizmo.com Accessed 21 July, 2010.
316
Syrdal, A. and Steele, S. (1985). Vowel F1 as a function of speaker fundamental
frequency. Journal of the Acoustical Society of America, 78, S56.
Thompson, W. C., & Schumann, E. L. (1987). Interpretation of statistical

evidence in criminal trials: the Prosecutor's fallacy and the Defence
attorney's fallacy. Law and Human Behaviour, 11 (3), pp. 167-187.
Traunmller, H. and Eriksson, A. (1995). The Frequency Range of the Voice

Fundamental in the Speech of Male and Female Adults (unpublished
manuscript) Visited August 9, 2013 from:
http://www2.ling.su.se/staff/hartmut/f0_m%26f.pdf.
Trouvain, J. (2004). Tempo variation in speech production. Implications for

speech synthesis. Doctoral dissertation published as Report Nr. 8 of Reports
in Phonetics, University of the Saarland.
Trudgill, P. (1981). Linguistic accommodation: Sociolinguistic observations on a

sociopsychological theory. In Hendrick, R., Mase, C. and Miller, M. F. (eds.),
Papers from the Parasession on Language and Behavior. Chicago, IL.: Chicago
Linguistic Society, pp. 193-247.
U.S. National Research Council (2009). Strengthening Forensic Science in the

United States: A Path Forward. Washington D.C.: The National Academies
Press.
van Leeuwen, D. A. and Brmmer, N. (2007). An introduction to application-

independent evaluation of speaker recognition systems. In Mller, C. (ed.)
Speaker Classification I, LNAI 4343. Berlin: Springer-Verlag, pp. 330-353.
Vermeulen, J. (2009). Beware of the Distance: Evaluation of Spectral

Measurements of Synthetic Vowels Recorded at Different Distances.
Unpublished; University of York. MSc.
Wardhaugh, R. (2006). An Introduction to Sociolinguistics (5th ed.). Oxford, UK:

Blackwell.
Williams, C. E. and Stevens, K. N. (1972). Emotions and speech: some acoustical

correlates. Journal of the Acoustical Society of America, 52, pp. 1238-1250.
Wright, M. (2005). Studies of the Phonetic-Interaction Interface: Clicks and

Interactional Structures in English Conversation. Unpublished; University of
York. PhD.
Wright, M. (2007). Clicks as markers of new sequences in English conversation.

Proceedings of the 16th International Congress of the Phonetic Sciences,
Saarbrcken (ICPhS XVI), pp. 1069-1072.
Wright, M. (2011a). On clicks in English talk-in-interaction. Journal of the

International Phonetic Association, 41(2), pp. 207-229.
317
Wright, M. (2011b). The phoneticsinteraction interface in the initiation of
closings in everyday English telephone calls. Journal of Pragmatics, 43(4), pp.
1080-1099.
Xue S. A. and Hao J. G. (2006). Normative Standards for vocal tract dimensions
by race as measured by acoustic pharyngometry. Journal of Voice, 20, pp.
391-400.
Zetterholm, E. (2006). Same speaker different voices: A study of one

impersonator and some of his different imitations. In Proceedings of the 11th
Speech Science and Technology Conference, Auckland, New Zealand, pp. 70-75.
Zhang, C., Morrison, G. S., Rose, P. (2008). Forensic speaker recognition in

Chinese: a multivariate likelihood ratio discrimination on /i/ and /y/.
Proceedings of Interspeech 2008: Incorporating SST 2008 and International
Speech Communication Association, pp. 19371940.
COURT CASES
Daubert v Merrell Dow Pharmaceuticals, Inc. (509 US 579 [1993]).
Frye v United States (293 F. 1013 D.C. Cir. [1923])
Joseph Crossfield and Sons Ltd v Techno Chemical Laboratories Ltd (29 TLR, 378
& 379 [1913])
Kumho Tire v Carmichael (526 U.S. 137 [1999])
R v Bonython [1984] 38 SASR 45
R v Doheny and Adams [1996] EWCA Crim 728
R v Flynn and St John [2008] EWCA Crim 970
R v Hufnagl [2008] NSWDC 134
R v Robb [1991] 93 EWCA Crim 161
R v T [2011] EWCA Crim 729
United States v Starzecpyzel (880 F. Supp. 1027 S.D.N.Y. [1995])
318

PHD GOLDerica

Uploaded by

Copyright:

Available Formats

PHD GOLDerica

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PHD GOLDerica

Uploaded by

Copyright:

Available Formats

Calculating likelihood ratios for

forensic speaker comparisons using

Erica Ashley Gold

The University of York

likelihood ratios using phonetic and linguistic parameters derived from a

corpus of recordings of speakers of Southern Standard British English. The

research serves as an investigation into the development of the numerical

likelihood ratio as a medium for framing forensic speaker comparison

conclusions. The thesis begins by investigating which parameters are claimed to

be the most useful speaker discriminants according to expert opinion, and in

turn examines four of these selected/valued parameters individually in

relation to intra- and inter-speaker variation, their capacities as speaker

parameters analyzed are articulation rate, fundamental frequency, long-term

formant distributions, and the incidence of clicks (velaric ingressive plosives).

under a numerical likelihood ratio framework in order to provide an overall

The contributions of this research are threefold. Firstly, the thesis

comparison practices around the world. Secondly, it expands the phonetic

literature by providing acoustic and auditory analysis, as well as population

have identified as effective speaker discriminants. And thirdly, it contributes to

by creating a human-based (auditory and acoustic-phonetic) forensic speaker

Table of Contents ......................................................................................................................................... 1

List of Tables and Figures .................................................................................................................... 10

List of Appendices .................................................................................................................................... 15

Quote Page .................................................................................................................................................... 21

1.1 Forensic Speaker Comparison .................................................................................................... 23

1.1.1 Expression of Conclusions ................................................................................................... 25

1.2 Research Aims ................................................................................................................................... 27

1.3 Thesis Outline .................................................................................................................................... 28

Chapter 2: Literature Review ............................................................................................................. 32

2.1 Paradigm Shift ................................................................................................................................... 32

2.2 Changes in the Law .......................................................................................................................... 35

2.3 Bayes Theorem................................................................................................................................. 38

2.3.1 Likelihood Ratio ....................................................................................................................... 40

2.3.3 Posterior Odds .......................................................................................................................... 44

2.3.4 Logical Fallacies ....................................................................................................................... 45

2.4 Forensic Speaker Comparison .................................................................................................... 47

2.4.1 Complexity of Speech Data .................................................................................................. 48

2.4.2 The Phonetic Shibboleth....................................................................................................... 51

2.4.2.1 Research Question 1 ..................................................................................................... 53

2.4.3 Current Conclusion Framework in the UK .................................................................... 53

2.5 Likelihood Ratios in Forensic Speech Science ...................................................................... 59

2.5.1 Likelihood Ratios in the Literature .................................................................................. 60

2.5.1.1 Likelihood Ratios for Speaker Discrimination ................................................... 60

2.5.1.1.1 Research Question 2.................................................................................. 61

2.5.1.2 Improving Likelihood Ratio Methodologies ....................................................... 62

2.5.1.2.1 Research Question 3 .................................................................................. 63

2.5.2 Likelihood Ratios in Practice .............................................................................................. 64

2.5.2.1 Research Question 4 ..................................................................................................... 68

2.6 Summary of Research Questions ............................................................................................... 69

Chapter 3: International Survey of Forensic Speaker Comparison Practices ......... 71

3.1.1 Background ................................................................................................................................ 72

3.2.2 Methodology: Data Compilation ........................................................................................ 74

3.3 Participants ......................................................................................................................................... 75

3.3.1 Countries ..................................................................................................................................... 75

3.3.2 Place of Work ............................................................................................................................ 76

3.3.3 Experience .................................................................................................................................. 76

3.4 Methods of Analysis ........................................................................................................................ 76

3.5 Conclusion Frameworks ................................................................................................................ 78