Singer Identification in Popular Music Recordings Using Voice Coding Features

Singer Identification in Popular Music Recordings

Using Voice Coding Features
Youngmoo E. Kim

Brian Whitman

MIT Media Lab

Cambridge, MA 02139
+01 617 253 0619
[email protected]

MIT Media Lab

Cambridge, MA 02139
+01 617 253 0112
[email protected]

ABSTRACT
In most popular music, the vocals sung by the lead singer are
the focal point of the song. The unique qualities of a singers
voice make it relatively easy for us to identify a song as
belonging to that particular artist. With little training, if one i s
familiar with a particular singers voice one can usually
recognize that voice in other pieces, even when hearing a song
for the first time. The research presented in this paper attempts
to automatically establish the identity of a singer using
acoustic features extracted from songs in a database of popular
music. As a first step, an untrained algorithm for automatically
extracting vocal segments from within songs is presented.
Once these vocal segments are identified, they are presented t o
a singer identification system that has been trained on data
taken from other songs by the same artists in the database.

1. INTRODUCTION
The singing voice is the oldest musical instrument and one
with which almost everyone has a great deal of familiarity.
Given the importance and usefulness of vocal communication,
it is not surprising that our auditory physiology and
perceptual apparatus has evolved to a high level of sensitivity
to the human voice. Once we are exposed to the sound of a
particular persons speaking voice, it is relatively easy t o
identify that voice, even with very little training. For the most
part the same holds true with regards to the singing voice.
Once we become familiar with the sound of a particular singers
voice, we can usually identify the voice, even when hearing a
piece for the first time.
Not only is the voice the oldest musical instrument, it is also
one of the most complex from an acoustic standpoint. This i s
primarily due to the rapid acoustic variation involved in the
singing process. In order to pronounce different words, a
singer must move their jaw, tongue, teeth, etc., changing the
shape and thus the acoustic properties of their vocal tract. No
other instrument exhibits the amount of physical variation of
the human voice. This complexity has affected research in both
analysis and synthesis of singing [1].
In spite of this complexity, voice identification is almost
effortless to us. But perhaps what is more remarkable is that
even in the presence of interfering sounds, such as instruments
or background noise, we can still identify the voice of a
familiar singer. Thus, our process of identification most likely
depends on features invariant to these environmental
variations. As will be discussed later, the search for such
invariant features that can be used for robust automatic
identification is no easy task.
Permission to make digital or hard copies of all or part of this
work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or
commercial advantage and that copies bear this notice and the
full citation on the first page.
2002 IRCAM Centre Pompidou

2. BACKGROUND
A significant amount of research has been performed o n
speaker (talker) identification from digitized speech for
applications such as verification of identity. These systems for
the most part use features similar to those used in speech
recognition. Many of these systems are trained on pristine data
(without background noise) and performance tends to degrade
in noisy environments. And since they are trained on spoken
data, they perform poorly to singing voice input. For more o n
talker identification systems, see [2].
In the realm of music information retrieval, there is a
burgeoning amount of interest and work on automatic song
and artist identification from acoustic data. Such systems
would obviously be useful for anyone attempting to ascertain
the title or performing artist of a new piece of music and could
also aid preference-based searches for music. Another area
where this research has generated a great deal of interest i s
copyright protection and enforcement. Most of these systems
utilize frequency domain features extracted from recordings,
which are then used to train a classifier built using one of
many machine learning techniques. Robust song identification
from acoustic parameters has proven to be very successful
(with accuracy greater than 99% in some cases) in identifying
songs included in the database [3]. Artist identification is a
much more difficult task, and not as well-defined as individual
song identification. A recent example of an artist identification
system is [4], which reports accuracies of approximately 50%
in artist identification on a database of about 250 songs.
Also relevant to the task of singer identification is work i n
musical instrument identification. Our ability to distinguish
different voices (even when singing or speaking the same
phrase) is akin to our ability to distinguish different
instruments (even when playing the same notes). Thus, it i s
likely that many of the features used in automatic instrument
identification systems will be useful for singer identification
as well. Work by Martin [5] on solo instrument identification
demonstrates the importance of both spectral and temporal
features and highlights the difficulty in building machine
listening systems that generalize beyond a limited set of
training conditions.
Obviously, singer identification and artist identification can
amount to the same thing in many situations. In [6],
Berenzweig and Ellis use vocal music as an input to a speech
recognition system, achieving a success rate of 80% i n
isolating vocal regions. In [7], Berenzweig, Ellis, and Lawrence
use a neural network trained on radio recordings to similarly
segment songs into vocal and non-vocal regions. By focusing
on voice regions alone, they were able to improve artist
identification by 15%.
The system presented here also attempts to perform
segmentation of vocal regions prior to singer identification.
After segmentation, the classifier uses features drawn from

Singer Identification in Popular Music Recordings Using Voice Coding Features

voice coding based on Linear Predictive Coding (LPC),
although with some modification. LPC is particularly good at
highlighting formant locations (regions of resonance, which
have been shown to be especially significant perceptually [8]).
Much research has been performed on speech coding via LPC,
and its uses are ubiquitous today. For example, all digital
cellular phones use LPC-based voice coders and maintain
fairly good sound quality, even at low information rates. Since
these models were designed to primarily carry the human voice
accurately, it seems logical that they could be useful for singer
identification as well.

3. DETECTION OF VOCAL REGIONS

Before trying to establish who is singing a particular piece, i t
is obviously important to identify the sections of that piece i n
which singing actually occurs. In this section, we present a
technique for automatically detecting these regions of singing
within recordings.

3.1 Vocal frequency regions

The majority of energy in the singing voice falls between 200
Hz and 2000 Hz (with some variation depending on the singer).
This is a region that the human auditory system is particularly
sensitive to. Classic experiments on a wide variety of human
listeners have established general equal-loudness curves [9]
which establish that signals within this mid-range of
frequencies are perceived as louder than signals of equivalent
absolute amplitude at other frequencies, higher or lower.
Another perceptual effect that predominates in this region i s
masking, in which energy in one frequency band will obscure
or mask lesser energies in adjacent frequency bands. In most
recorded vocal music, the tendency is to isolate other
instruments away from the voice so as not to mask or be
masked by the voice. One notable exception is singing with a
symphony orchestra. Many orchestral instruments fall in the
same frequency range as the voice, and the sheer number of
instruments is more than enough to mask the untrained voice.
Classically trained singers, however, are taught to produce
extra resonance at a higher frequency (often referred to as the
singers formant, located around 2500 Hz), which allows the
voice to be perceived against the overwhelming numbers [10].
But since the majority of popular recorded music is not in this
style, we can restrict our range of interest to lower frequencies.
Since we are interested in detecting regions of singing, a
straightforward method would be to detect energy within the
frequencies bounded by the range of vocal energy. A very
simple approach is to filter the audio signal with a band-pass
filter which allows the vocal range to pass through while
attenuating other frequency regions. For this, we use a simple
Chebychev infinite-impulse response (IIR) digital filter of
order 12. The frequency response of this filter is shown i n
Figure 1.

for example, disperse energy over a wide range of frequencies, a

significant amount of which falls in our range of interest. So
another measure is needed to discriminate the voice from these
other sources.

3.2 Detection via harmonicity

Singing primarily consists of sounds generated by phonation,
the rapid vibration of the vocal folds resulting in utterances
referred to as voiced by speech researchers. This is as opposed
to unvoiced sounds which are generated by the turbulence of
air against the lips or tongue, such as the consonants [f] or [s].
Singing is >90% voiced, whereas speech is only ~60% voiced
[11]. Because of this, the singing voice is highly harmonic
(energy exists at integer multiples of the fundamental
frequency, or pitch). Other high energy sounds in this region,
drums in particular, are not as harmonic and distribute their
energy more widely in frequency.
To exploit this difference, we use an inverse comb filterbank t o
detect high amounts of harmonic energy. The block diagram
and frequency response of a simple inverse comb filter i s
shown in Figure 2. By passing the previously filtered signal
(as described above) through a bank of inverse comb filters
with varying delays, we can find the fundamental frequency
which the signal is most attenuated. By taking the ratio of the
total signal energy to the maximally harmonically attenuated
signal, we have a measure of harmonicity, or how harmonic the
signal is within the analysis frame.

H=

E original

(1)

min( E filtered ,i )
i

By thresholding the harmonicity against a fixed value, we have

a detector for harmonic sounds. The hypothesis is that most of
these correspond to regions of singing. Results using this
technique are presented in Section 5.
x[n]

y[n]

z -N

Figure 2: Block diagram of the simple inverse comb filter

(top). The frequency response of an inverse comb filter, tuned
to ~400 Hz (bottom). The spacing of the attenuated
frequencies is determined by the delay parameter N.

4. SINGER IDENTIFICATION

Figure 1: Vocal enhancement filter frequency response.

This filter has the musical effect of attenuating other
instruments that fall outside of this frequency region, such as
bass and cymbals. But even in popular music, the voice is not
the only instrument producing energy in this region. Drums,

Although relatively easy for humans, robust singer

identification is an extremely difficult task for a machine
listening system. Even with clean signals (with no other
instruments or background noise), simple frequency or time
domain features do not lead easily to a unique voiceprint.
And in most performances or recordings where the voice i s
amidst a mixture of other sounds, the problem becomes even
more complex. This section discusses the features, extraction
methods, and classification techniques used in the singer ID
system.

Singer Identification in Popular Music Recordings Using Voice Coding Features

4.1 Features from Speech Coding
Much research (primarily dealing with speech) has focused o n
voice coding using an analysis/synthesis approach. In this
approach a source signal is analyzed and re-synthesized
according to a source-filter model of the human voice. This i s
the general principle behind Linear Predictive Coding (LPC).
The primary advantage of this technique has been its utility i n
compressing speech data resulting in the low-bitrate speech
coders used in many applications today.

4.1.1 Traditional LPC

The goal of linear predictive analysis is to establish an

estimate, s[n ] to the source signal s [n ] , using a linear
combination of p past samples of the input signal:
p

s[n ] =

a s[n - k ]

(2)

k =1

The coefficient values a k in Equation (2) are determined by

minimizing the mean squared prediction error, which is the
difference between the source signal and the predicted signal:
Em =

(s [n] - s [n])
m

(3)

The transfer function relating the source signal and the signal
estimate is shown [12] to be an all-pole filter:
G
H[ z ] =
(4)
A[z ]
where the denominator is defined as follows:
p

A[z ] = 1 -

a z
k

-k

(5)

k =1

This demonstrates how linear predictive analysis is equivalent

to a source-filter model, where the vocal tract response i s
modeled using a time-varying all-pole filter function of order
p. The calculated coefficients can be factored to determine the
pole locations, which generally correspond to the formants
(resonances) of the vocal tract. An example of an LPC filter
response is shown in Figure 3.

is also believed that an individuals particular formant

patterns are a key feature for speaker identification [8]. Due t o
increased harmonic energy in singing, formants are enhanced,
so it is reasonable to believe that they would be important for
singer identification as well. Formant frequencies and
magnitudes extracted via LPC (with 12 poles) are used as key
features in the singer classifier.
A common technique for minimizing the prediction error
(Equation 3) uses the autocorrelation matrix (hence its name,
the autocorrelation method), which can be calculated from the
power spectrum of the signal. This will be useful in the next
section, which discusses warping of the power spectrum t o
better fit established theory on auditory perception.

4.1.2 Warped Linear Prediction

One disadvantage of standard LPC is that it treats all

frequencies equally on a linear scale. However, the human ear i s
not equally sensitive to all frequencies linearly. In fact our
frequency sensitivity is very close to logarithmic. As a result,
LPC systems sometimes place poles at higher frequencies
where the ear is less sensitive and miss closely spaced resonant
peaks at lower frequencies where the ear is more sensitive.
Using a higher order LPC is one way of compensating for this,
but increasing the number of poles makes it difficult to track
correlations between analysis frames.
Instead, we use a Warped Linear Prediction model by prewarping the power spectrum of each frame [13], [14]. The
warping function can be made to closely approximate the Bark
scale, which approximates the frequency sensitivity of human
hearing. Warping is achieved through the following relation:

asinw
w = w + 2tan-1

1- acos w

(6)

A parameter value of a=0.47 closely matches the Bark scale.

This allows us to keep the order of the analysis low (and thus
more easily track formants from frame to frame) while more
accurately capturing formant locations, especially at lower
frequencies. The additional emphasis on lower frequencies has
the added benefit of being able to pick out individual low
harmonics. Standard LPC does not have the resolution at low
frequencies to detect individual harmonics, but warped LPC i s
oftentimes able to identify the lowest harmonic, which usually
corresponds to the pitch, which can be a useful feature i n
singer identification. Figure 4 shows an example of warped LP.

Figure 3: A vocal segment (top) and its spectrum (bottom).

The line in the bottom figure shows the 12-pole LPC estimate.
The peaks indicate the formant locations.
As regions of high energy within a frequency range of
enhanced sensitivity, vocal formants are of special perceptual
significance. The general pattern of formant locations
determines our perception of phonemes, and thus language. It

Figure 4: The same vocal segment analyzed via warped

Linear Prediction. The warped frequency scale is on top and
the unwarped scale on the bottom. Note the particularly good
fit at lower frequencies.

Singer Identification in Popular Music Recordings Using Voice Coding Features

4.2 Classification Techniques
Two different classifiers were trained using established pattern
recognition algorithms. A brief description of the two
classifiers implemented follows. In this task, each class
represents an individual singer.

4.2.1 Gaussian Mixture Model (GMM)

The Gaussian mixture model uses multiple weighted Gaussians

to attempt to capture the behavior of each class of training
data. The use of multiple Gaussians is particularly beneficial
when analyzing data that has a distribution not well modeled
by a single cluster. It is a very flexible model that can adapt t o
encompass almost any distribution of data. Test points are
classified by a maximum likelihood discriminant function,
calculated by their distances from the multiple Gaussians of
the class distributions [15].
To determine the parameters of the Gaussians that best model
each class, we use the well-known technique of Expectation
Maximization (EM). EM is an iterative algorithm that
converges on parameters that are locally optimal according t o
the log-likelihood function. Thus, it is sensitive to initial
conditions, and usually several runs are performed to ensure
that the derived fit is a relatively good one. It is also useful t o
perform Principle Components Analysis (PCA) prior to EM.
PCA is a multi-dimensional rotation of the data onto the axes
of maximal variance. It also has the added benefit of
normalizing the data variances, which avoids highly different
scaling among the dimensions, which is problematic for EM.
The number of Gaussians used is generally a user-defined
parameter, and some experimentation is usually required t o
find a reasonable number for a given data set.

4.2.2 Support Vector Machine

A Support Vector Machine (SVM) is based on statistical error

minimization techniques applied to a machine learning
domain [16]. SVMs work by computing an optimal hyperplane
that can linearly separate (in one case) two classes of data.
These hyperplanes simplify to a set of Lagrange multipliers for
each training case, and the set of points within the dimensional
vectors fed for training that have non-zero Lagrangians are the
support vectors. The machine saves these support vectors and
applies them to new data in the form of the test set for further
on-line classification.
We used an SVM with a Gaussian kernel. Our C (maximum
lagrangian value) was set to 10. Each class was trained as a
separate SVM with all the positive examples from that class
along with the same amount of randomly chosen negative
examples. After training each individual SVM, we applied the
confidence thresholding metric discussed in [4] to remove
uncertain frame classifications and chose the SVM with the
highest confidence to classify frames in the test set [17].

5. EXPERIMENT AND RESULTS

Several different experiments were conducted using the
features and algorithms described in sections above. First, the
data set used for training and testing is described. Then the
specifics regarding the experimental configuration and
methods are presented, followed by the reporting of
experimental results.

5.1 The Data Set

The data sets used in this experiment are various subsets of the
NECI Minnowmatch testbed [4] with some minor additions.
The entire Minnowmatch testbed consists of >250 songs from
the albums of more than 20 distinct artists/groups. Some of
these songs, however, do not contain singing, and were
eliminated from consideration. Additionally, on occasion

different individuals sing some of the songs performed by a

single group. Care was taken to classify each song by the
actual singer on the recording. After these considerations, the
resulting testbed included 17 different solo singers and
slightly more than 200 songs. All songs were downsampled t o
11.025 kHz (from the CD sampling rate of 44.1 kHz) to reduce
the data storage and processing requirements. Even at the lower
sampling rate, most vocal energy falls well below the Nyquist
rate (half the sampling-rate) and is preserved for our analysis.

5.2 Experimental Procedure and Results

5.2.1 Detection of Vocal Regions

To test the accuracy of the vocal segment detector, a subset (20

songs, or approximately 10%) of the database was segmented
manually by ear into regions of singing and non-singing t o
provide a set of ground truth data. The relatively small size
of the ground truth data set is because of the tedious and
somewhat ill-defined nature of this task. Establishing exactly
where a vocal segment begins and ends with certainty i s
problematic. Low-level background vocals that tend to fade i n
and out in some songs add further complications. The
segmentation on this set was as accurate as can be, given the
difficulties.
This set of 20 songs was then analyzed using the vocal
detection system described in Section 3. Analysis was
performed using frame size of 1024 samples (~100 msec) and
frames were taken every 512 samples (~50 msec). The initial
threshold for vocal classification was a harmoncity value of
H>2. As can be seen from the results in Table 1, the classifier i s
not very accurate, though it does perform better than chance
(two classes = 50%). As a front-end for the singer identification
system, however, we would like to avoid false positives
(identifying regions of non-vocal music as having vocals),
since the identity classifier would attempt to place these nonvocal segments with a singer anyway. On the other hand, false
negatives (classifying regions of vocals as being instrumental
only) are more acceptable since they simply reduce the amount
of data fed to the singer ID system. Given this tradeoff, we can
greatly reduce the number of false positives by raising the
threshold of H , at the expense of more false negatives. As
shown in Table 3, raising the threshold to H=2.6 reduced the
error rate among non-vocal frames to ~20% while retaining
~30% of the vocal frames. This value was used to automatically
segment the data in the singer identification system.
Table 1: Performance of vocal detector at multiple
harmonicity thresholds.
H
Threshold

Vocal
Segments

Non-vocal
Segments

All
Segments

2.0

55.4%

53.1%

55.4%

2.3

40.5%

69.2%

55.1%

2.6

30.7%

79.3%

54.9%

5.2.2 Singer Identification

For singer classification, we used approximately half of the

database (the odd numbered songs from albums) to train the
classifier and the remaining songs to evaluate the performance
of our classifier. We conducted two sets of experiments: In the
first set, LPC features were extracted from entire songs and
used for classification. The second set of experiments used
only features from regions classified as containing vocals. In
both experiments, analysis frames were again 1024 samples
calculated at 512 sample intervals. A 12-pole LP analysis was
performed on both linear and warped scales. The frequencies
and magnitudes of the pole locations were used as inputs t o
the classifiers.

Singer Identification in Popular Music Recordings Using Voice Coding Features

Three different feature sets (linear scale data, warped scale data,
and both linear and warped data) were tested, and two different
classifiers (GMM and SVM) were used in each case. In the SVM
classifier, only every tenth data frame was used because of
computer memory constraints. The results from these
experiments are summarized in Table 2. The highest
performing number of Gaussians was used as the GMM result.
Table 2: Singer classification results. Results are listed as
percentages of songs correctly classified (followed by
percentages of individual frames correctly classified).
Experiment 1: Entire song data
Features

GMM

SVM

Linear frequency features

32.1 (16.6)

39.6 (30.7)

Warped frequency features

31.3 (17.1)

35.0 (30.4)

Linear and warped features

33.4 (16.5)

45.3 (29.6)

Experiment 2: Only song segments classified as vocals

Features

GMM

SVM

Linear frequency features

36.7 (18.1)

35.8 (17.6)

Warped frequency features

33.0 (17.4)

34.0 (26.8)

Linear and warped features

38.5 (16.6)

41.5 (28.8)

On the whole, the classification results are far greater than

chance (17 classes = ~6%), but still fall well short of expected
human performance. In general, the linear frequency features
tend to outperform the warped frequency features when each i s
used alone, but using them together does benefit performance.
Strangely, song and frame accuracy increases when using only
vocal segments in the GMM, but decreases in the SVM using
the same segments.

The singer classification results are notable in a few ways. Both

classification techniques found something discriminatory
within the features. With the GMM, accuracy increased as
expected when the classifier was trained and tested with the
voice-classified frames. That the performance of the SVM
decreased is a bit puzzling. It is likely that the SVM is finding
aspects of the features that are not specifically related to the
voice. The fewer number of frames in the second experiment
might account for the decrease in accuracy. There is also a great
deal of uncertainty about the accuracy of the voice labeling,
and that is likely factor as well.
The better performance of the linear frequency scale features vs.
the warped frequency features probably indicates that the
machine finds the increased accuracy of the linear scale at
higher frequencies useful. Though this is contrary to human
auditory perception, it is not surprising that there i s
discriminatory information there, though it is probably not
correlated with the voice. The increased performance in using
both linear and warped features indicates that the analyses are
not completely redundant.
Given the relatively low frame accuracy reported in the singer
identification experiments, the overall frame confusion matrix
(Figure 5) is not surprising. There is not a high amount of
intensity along the diagonal, except for a few particular artists.
It is as yet unclear why these particular singers are easily
detected. There may have been particular qualities to these
voices or other parts of the songs, which may have been
highlighted by the linear predictive analysis. This remains an
ongoing investigation in our research.

6. DISCUSSION AND FUTURE WORK

As shown in the results, the raw accuracy of the vocal detector
could use some improvement. But given the uncertainties i n
the defining the exact boundaries of vocal and instrumental
segments, the raw output numbers have some uncertainty
attached as well. It should be noted that the singing detector
presented here is an untrained system (it possesses no prior
knowledge). Other systems ([6] and [7]) have achieved higher
vocal detection accuracy (as well as [18] for speech vs. music),
but have been trained on ground-truth databases. It would be
possible to combine features from both systems to achieve
greater accuracy. A better perceptual model (than the static
filter used here) may be of substantial benefit as well. We are
currently investigating these possible improvements.
Qualitative listening to the detected regions demonstrates
some interesting points. Many of the extended mislabeled
regions occur at the beginning of songs or during instrumental
bridges where producers have highlighted instruments in the
absence of vocals. Also, the beginnings and ends of phrases
tend to be cut off, since they contain less harmonic content.
Regions of extended vowels (such as held notes) are
particularly well detected. A simple extension to the system
would be to pad each section with extra time on either side. But
whether this would aid in singer identification is an entirely
different question. It is conceivable that higher-level musical
knowledge could be added to the system in an attempt t o
identify song structure, such as the location of verses and
choruses, from patterns in the segmentation data. The
probability of vocals in those sections could be weighted more
strongly than in others, which would reduce the problem of
falsely classifying strong solo instrumental passages as
vocals.

Figure 5: Confusion matrix for all voiced data frames, using

both linear and warped features.
Others (e.g. [4] and [7]) have achieved higher artist
classification accuracy on the same test data, but have not
attempted to identify the individual singers. In [7], using only
vocal segments improved accuracy in a neural net classifier
using Mel-frequency Cepstral Coefficients (MFCCs). But it is
unclear in that study (and in the research presented here as
well) whether the classifier is actually training on vocal
features or is using some other aspect of the recordings, even
though both MFCCs and linear prediction have proven useful
in speech applications. Their system also uses the differences
between MFCCs as features, indicating that using differential
magnitudes (in addition to or instead of raw magnitudes) may
be beneficial.
Another possibility for improving the singer classifier is t o
incorporate more time-varying information. While almost all
audio analysis is conducted at a fixed rate, the actual
information rate of audio varies widely. For this reason, a statebased classifier, such as a Hidden Markov Model, may improve

Singer Identification in Popular Music Recordings Using Voice Coding Features

classifier performance. This will be explored in our research at
a later date.

7. ACKNOWLEDGEMENTS

[8] R. Brown, An experimental study of the relative

importance of acoustic parameters for auditory speaker
recognition, Language and Speech, vol. 24, pp. 295-310,
1981.

Our thanks to Dr. Paris Smaragdis for his help in enlightening

us to some of the more obscure aspects of pattern recognition.
And thanks to Prof. Barry Vercoe for his support of our
research.

[9] H. Fletcher, Auditory patterns, Review of

Physics, vol. 12, pp. 47-65, 1940.

8. REFERENCES

[11] P. R. Cook, Identification of Control Parameters in a n

Articulatory Vocal Tract Model, with Applications to the
Synthesis of Singing. Ph.D. Thesis. Stanford University,
Stanford, CA, 1990.

[1] Y. E. Kim, Excitation Codebook Design for Coding of the

Singing Voice, Submitted to the 2001 IEEE Workshop o n
the Applications of Signal Processing to Audio and
Acoustics, 2001.
[2] R. Mammone, X. Zhang, and R. P. Ramachandran, Robust
speaker recognition: A feature-based approach, IEEE
Signal Processing Magazine, vol. 13, pp. 58-71, 1996.
[3] J. Herre, E. Allamanche, and O. Hullmuth, Robust
Matching of Aduio Signals Using Spectreal Flatness
Features, presented at IEEE Workshop on Applications o f
Signal Processing to Audio and Acoustics, New Paltz, NY,
2001.
[4] B. Whitman, G. Flake, and S. Lawrence, Artist detection i n
music with Minnowmatch, presented at IEEE Workshop
on Neural Networks for Signal Processing, Falmouth,
MA, 2001.
[5] K. Martin, Sound-Source Recognition: A Theory and
Computational Model. Ph.D. Thesis. Massachusetts
Institute of Technology, Cambridge, MA, 1999.
[6] A. L. Berenzweig and D. P. W. Ellis, Locating Singing
Voice Segments Within Music Signals, presented at IEEE
Workshop on Applications of Signal Processing to Audio
and Acoustics, New Paltz, NY, 2001.
[7] A. Berenzweig, D. P. W. Ellis, and S. Lawrence, Using
Voice Segments to Improve Artist Classification of
Music, In press, 2002.

Modern

[10] J. Sundberg, The Science of the Singing Voice. Dekalb, IL:

Northern Illinois University Press, 1987.

[12] L. R. Rabiner and R. W. Schafer, Digital Processing o f

Speech Signals. Englewood Cliffs, NJ: Prentice-Hall,
1978.
[13] H. W. Strube, Linear prediction on a warped frequency
scale, Journal of the Acoustical Society of America, vol.
68, pp. 1071-1076, 1980.
[14] A. Hrm, A Comparison of Warped and Conventional
Linear Predictive Coding, IEEE Transactions on Speech
and Audio Processing, vol. 9, pp. 579-588, 2001.
[15] R. Duda, P. Hart, and D. Stork, Pattern Classification, 2nd
ed. New York: John Wiley & Sons, 2000.
[16] C. J. C. Burges, A tutorial on support vector machines for
pattern recognition, Data Mining and Knowledge
Discovery, vol. 2, pp. 955-974, 1998.
[17] G. Flake, NODElib,.: NEC Research Institute.
[18] E. D. Scheirer and M. Slaney, Construction and
evaluation of a robust multifeature speech/music
discriminator, presented at IEEE Conference on
Acoustics, Speech, and Signal Processing (ICASSP),
Munich, Germany, 1997.