Bangla Text to Speech using Festival

Firoj Alam S.M. Murtoza Habib Mumit Khan

Center for Research on Ban- Center for Research on Ban- Center for Research on Ban-
gla Language Processing gla Language Processing gla Language Processing and
BRAC University. BRAC University. Department of Computer
[email protected] [email protected] Science and Engineering
BRAC University.
[email protected]

ternet and enhancing other information systems.

Abstract A touch screen based kiosk that integrates a
Bangla TTS has the potential to empower the
This paper describes the development of the 49% of the population who are illiterate 3 . A
first, usable, open source and freely available screen reader that integrates a Bangla TTS will
Bangla Text to Speech (TTS) system for Ban- do the same for the estimated 100 thousand vi-
gladeshi Bangla using the open source Festival sually impaired citizens of Bangladesh. A Text to
TTS engine. Besides that, this paper also dis-
Speech is a computer based system capable of
cusses a few practical applications that use this
system. This system is developed using di- converting computer readable text into speech.
phone concatenation approach in its waveform There are two main components such as Natural
generation phase. Construction of a diphone Language Processing (NLP) and Digital Signal
database and implementation of the natural Processing (DSP) [Thierry 1997][Paul
language processing modules are described. 2009][A.W. Black et el. 2003]. The NLP com-
Natural language processing modules include ponent includes pre-processing, sentence split-
text processing, tokenizing and grapheme to ting, tokenization, text analysis, homograph reso-
phoneme (G2P) conversion that were written lution, parsing, pronunciation, stress, syllabifica-
in Festival's format. Finally, a test was con- tion and prosody prediction. Working with pro-
ducted to evaluate the intelligibility of the syn-
nunciation, stress, syllabification and prosody
thesized speech.
prediction sometime is termed as linguistic anal-
Index Terms: speech synthesis, diphone ysis. Whereas, the DSP component includes
segment list generation, speech decoding, proso-
1 Introduction dy matching, segment concatenation and signal
synthesis. Pre-processing is the process of identi-
Bangla (exonym: Bengali) is one of the most fying the text genre, character encoding issues
widely spoken languages of the world (it is and multilingual issues. Sentence splitting is the
ranked between four 1 and seven 2 based on the process of segmenting the document text into a
number of speakers), with nearly 200 million list of sentences. Segmenting each sentence into
native speakers. However, this is one of the most a list of possible tokens can be done by tokeniza-
under-resourced languages which lack speech tion. In text analysis part, different semiotic
applications. The aim of this project is to develop classes were identified, and then using a parser
a freely available Bangla text to speech system. each token is assigned to a specific semiotic
A freely available and open-source TTS system class. After that, verbalization is performed on
for Bangla language can greatly aid human- non-natural language token. Homograph resolu-
computer interaction: the possibilities are endless tion is the process of identifying the correct un-
– such a system can help overcome the literacy derlying word for ambiguous token. The process
barrier of the common masses, empower the vi- of generating pronunciation from orthographic
sually impaired population, increase the possibil- representation can be done by pronunciation lex-
ities of improved man-machine interaction icon and grapheme-to-phoneme (G2P) algorithm.
through on-line newspaper reading from the in- Prosody prediction is the process of identifying
the phrase break, prominence and intonation
1
http://www2.ignatius.edu/faculty/turner/languages.htm,
Last accessed December 26, 2007.
2 3
http://en.wikipedia.org/wiki/List_of_languages_by_total_s Bangladesh Bureau of Statistics, 2004.
peakers, Last accessed December 26, 2007. http://www.bbs.gov.bd/dataindex/stat_bangladesh.pdf

154

Conference on Human Language Technology for Development, Alexandria, Egypt, 2-5 May 2011.
tune. There has not been much work done on The presence and absence of heatstroke has sig-
prosody in this paper. DSP component or wave- nificant implications to distinguish consonant
form generation is the final stage of a TTS. This conjunct, dependent and independent vowel.
involves the production of acoustic signals using Words are delimited by a space in general. Vo-
a particular synthesis approaches such as formant wels have corresponding full-character forms
synthesis, articulatory synthesis and concatena- when they appear in an absolute initial position
tion based synthesis. The attempt that has been of a word. Generally a vowel followed by a con-
made here is the second generation diphone con- sonant takes a modified shape and placed at the
catenation based synthesis, using widely usable left, right or both, or at the bottom of the conso-
Festival framework [A.W. Black et al. 2003]. nant which are signifies as vowel modifiers. The
inventory of Bengali script is made up of 11 vo-
2 Literature survey wels, 39 consonants, 216 consonant conjuncts,
10 digits, modifiers, punctuation marks and a
Significant effort has been made for different
few symbols [Bangal Academy 1992][Wikipedia
languages to develop TTS using the Festival
2010]. The vowel and consonant characters are
framework such as English, Japanese [A.W.
called basic characters. The consonant conjunct
Black et al. 2003], Welsh [R.J. Jones et al. 2006],
is joined by ‘hasanta’. The concept of upper and
Telugu [C. Kamisetty et al. 2006], [S.P. Kishore
lower case is absent in Bangla script. English
et al. 2002], Hindi [S.P. Kishore et al. 2002],
equivalent full stop is the Bengali punctuation
[A.G. Ramakishnan et al. 2004], Turkish [Ö. Sa-
lor et al. 2003] and Sinhala [Ruvan et al. 2007]. mark the down-stroke dari (।); Unicode - \u0964.
However, very little work has been done on Commas, semicolons, colons, quotation marks,
Bangla. Several attempts have been made in the etc. are the same as in English.
past, where different aspects of a Bangla TTS
system were covered in [Tanuja et al. 2005], 3.2 Bangla phoneme inventory
[Asok 2002], [Shyamal et al. 2002] and [Ani- The phoneme inventory of Bangla consists of 30
ruddha et al. 2004]. [Tanuja et al. 2005] showed consonants, 14 monophthong vowels (oral and
different TTS modules such as optimal text se- nasal vowels) and 21 diphthongs [Firoj et al.
lection, G2P conversion and automatic segmen- 2008 (b)] [Firoj et al. 2008 (a)]. Consonants and
tation with experimental results. Phoneme and vowels are shown in Table 1 and Table 2. The
partname were used to develop voice database diphthongs are the following: অও/ɔo/, আই/ai/, আউ
and ESNOLA technique were used for concate-
nation [Asok 2002], but the quality of the system /au/, আ /aja/, ইউ/iu/, ইএ-ই /ie/, ইও/io/, ই -ইআ
suffers due to the lack of smoothness. Shyamal et /ia/, উই/ui/, উ -উআ /ua/, উ /ue/, উ -উও/uo/,
al. [Shyamal et al. 2002] showed some practical এই/ei/, এউ/eu/, এও/eo/, এ -এআ /ea/, এ /æa/,
applications with Bangla TTS system using ES-
ওই/oi/, ওউ/ou/, ও -ওআ /oa/, ও /oe/.
NOLA technique. In [Aniruddha et al. 2004] au-
thor showed the pronunciation rule and phoneme
to speech synthesizer using formant synthesis Front Central Back
technique. Another attempt has been made to High ই/i, ইঁ , উ/u/, উঁ /
develop Bangla TTS using multisyn unit selec-
tion and unit selection technique within Festival High-Mid এ/e, এঁ ও/o, ওঁ
framework in [Firoj et al. 2007] but the system Mid-Low এযা ӕ,এযাঁ ӕ অ/ɔ/, অঁ/ɔ/
was developed on a limited domain and could
not be used as a general purpose TTS system. To Low আ/a, আঁ
the best of our knowledge this is the first com-
plete work for general purpose, open source,
freely available and platform independent Bangla Table 1: Vowel phoneme inventory
Text to Speech system for Bangladeshi Bangla. 3.3 Natural language processing in Festival
3 Development Festvox [A.W. Black et al. 2003] provides dif-
ferent natural language processing modules for
3.1 Bangla writing system building a new voice. These modules can be
Bangla is written left to right in horizontal lines generated automatically which appears as a form
with a left-to-right heatstroke (is called matra). of scheme files. The scheme files need to be cus-
tomized for a new language. The language spe-

155
cific scripts (phone, lexicon and tokenization) whitespace and the punctuation marks, which is
and speaker specific scripts (duration and intona- used in our implementation to tokenize Bangla
tion) can be externally configured and imple- text. After tokenization, text normalization is
mented without recompiling the system [A.W. performed. In text normalization, the first task is
Black et al. 2003]. Since the templates are to identify the semiotic classes. The following
scheme files, which is typically an interpreted section discusses the semiotic class [Paul 2009]
language, so recompilation is not required. The (as opposed to say NSW) identification, tokeni-
following NLP related tasks are involved when zation and standard word generation and disam-
building a new voice in Festvox: biguation rule. Moreover, this work has been
• Defining the phoneset done separately before implementing into Festiv-
• Tokenization and text normalization al.
• Pronunciation: Lexicon and grapheme to pho- We identified a set of semiotic classes which be-
neme conversion. long to the Bangla language. To do this, we have
• Implementation of syllabification and stress selected a news corpus [Prothom-Alo 2009]
• Prosody: Phrase breaking, accent prediction, [Khair et al. 2006] with 18100378 tokens and
assignment of duration to phones and generation 384048 token types [Khair et al. 2006], forum
of f0 contour. [forum.amaderprojukti.com 2008] and blog
[www.somewhereinblog.net 2008]. After that we
Bilabial Dental Alveolar Post- Alveo- Palatal Velar Glottal
lar
voiceless ঩ /p/ প /ph/ ত / থ h
/ ট /t/ ঠ /th/ চ /c/ ছ /ch/ ক /k/ খ /kh/
Stops

voiced ফ /b/ ব /b /
h
দ / ধ h
/ ড /d/ ঢ /d /
h
ম, জ/ɟ/ ঝ /ɟ /
h
গ /g/ ঘ /gh/

Nasals ভ/m/ ন,ণ /n/ ঙ,ং /ŋ/

Trill য /r/
Flap ড়, ঢ় /ɾ/
Fricatives ঱,঳ /s/ ঱,ল,঳ /ʃ/ ঴,ং /h/
Lateral র /l/
Approximant ৞ /j/
Table 2: Consonant phoneme inventory
proceeded in two steps to identify the semiotic
3.3.1 Defining phoneset for Bangla classes. Firstly, a python [Python 2008] script
The phoneset that has been explained in section was used to identify the semiotic class from news
3.2 was implemented in festival format which corpus and manually checked the semiotic
had to be transcribed into ASCII. Festvox has a classes in the corpus of forum and blog. Second-
separate module to implement this phoneset with ly, we defined a set of rules according to context
their features. For vowel, the features include of homographs or ambiguous tokens to find the
height, length (e.g: short, long, diphthong and semiotic classes. The resulted set of semiotic
schwa), front vs back, lip rounding, and tense vs classes of Bangla text is shown in Table 3.
lex. For consonant, the features include place of Semiotic Example
articulation (e.g: bilabial, dental, alveolar, post- class/token
alveolar, palatal, velar and glottal), manner of type
articulation (e.g: stop, nasal, tril, flap, fricative, English text জাবা Platform Independent ফলর
lateral, glide/approximant), aspiration and voic- Bangla text আ
ing. Numbers ১২১,২৩,২৩৪; ১ভ, ২৞, ৩৞; I, II, III,
(cardinal, ১২.২৩, ২৩,৩৩.৩৩; ১/২, ২৩/২৩; ১২:১২
3.3.2 Text analysis and text normalization ordinal, roman,
Like any conventional writing system, Bangla floating
number,
script uses whitespace as a delimiter which fraction, ratio)
helped us to make a one-to-one mapping be- Telephone and ০২৯৫৬৭৪৪৭; ০১৫২৩০৩৩৯৮ (19
tween tokens and words. Besides whitespace, mobile number different formats)
Bangla script also uses punctuation (such as dari-
Years ২০০৬; ১৯৯৮; ৯৮ ঳ালর
।, ?, !, ;) as a delimiter. The default text tokeniza-
Date ০২-০৬-২০০৬ (12 different formats)
tion methodology available in Festival is the

156
 Time ৪.২০ মভ ; ৪.২০ মভমনট; Context dependent hand written rules were ap-
plied for these ambiguities. In case of Bangla,
Percentage ১২%
after time pattern ১২. ৩০ (12.30) we have a to-
Money ১০ ৳
ken (minute), so we look at the next token
E-mail আভায ই-মভইর মঠকানা: and decide whether it is time or a floating point
[email protected]
number. There are rare cases where context de-
URL ঳পটও৞যাযমট http://googlecode.com pendent rules fail in year-number ambiguity then
঳াইট we verbalize the token as a pair of two digits. For
Abbreviation ড ; মভা ; ঳া example, the token ১৯৯৮ (1998), we expand it as
Acronym ঢামফ; ফাউমফ, মকমফ উ আ ই (Nineteen hundred ninety
Mathematical (১+২=৩) eight) rather than এ আ ই
equation (one thousand nine hundred ninety eight). The
natural language text is relatively straightfor-
Table 3: Possible token type in Bangla text ward, and Bangla does not have upper and lower
case. The system implemented based on the work
A set of tags defined for each semiotic class and of [Firoj et al. 2009], claims that the accuracy of
assigned these tags to each class of tokens. The the ambiguous token is 87%.
tokenization undergoes three levels such as: i.
Tokenizer ii. Splitter and iii. Classifier. Whites- 3.3.3 Pronunciation
pace is used to tokenize a string of characters This system takes the word based on orthograph-
into a separate token. Punctuations and delimi- ic linguistic representation and generates a pho-
ters were identified and used by the splitter to nemic or phonetic description of what is to be
classify the token. Context sensitive rules were spoken by the subsequent phases of TTS. In ge-
written as whitespace is not a valid delimiter for nerating this representation we used a lexicon of
tokenizing phone numbers, year, time and float- known words and a grapheme-to-phoneme (G2P)
ing point numbers. Finally, the classifier classi- algorithm to handle proper names and unknown
fies the token by looking at the contextual rule. words.
For each type of token, regular expression were We developed a system lexicon [2, pp215]
written in festival scheme. where the entries contain orthography and pro-
The token expander expands the token by verba- nunciation in IPA. Due to the lack of a digitized
lizing and disambiguating the ambiguous token. offline lexicon for Bangla we had to develop it
Verbalization [Paul 2009] or standard word gen- manually by linguistic experts. To the best of our
eration is the process of converting non-natural knowledge this is the first digitized IPA incorpo-
language text into standard words or natural lan- rated and syllabified lexicon. The lexicon con-
guage text. A template based approach [Paul tains 93K entries where 80K entries entered by
2009] such as the lexicon was used for number hand and the rest of them were automatically
cardinal, ordinal, acronym, and abbreviations. generated by G2P system [Ayesha et al. 2006].
Abbreviations are productive and a new one may The performance of this G2P system is 89.48%.
appear, so an automatic process may require Therefore, the automatically generated entries
solving unknown abbreviations. In case of Ban- had to be checked manually to maintain the qual-
gla acronyms, most of the time people say the ity of the lexicon by expert linguists. The system
acronym as a full form without expanding it. For is now available in online for public access
example, ɔk/ expands to [CRBLP 2010]. Another case needs to be han-
ɔmon komiʃɔn/ but people say it dled in order to implement the lexicon into Fes-
tival. The Unicode encoded phonetic representa-
as ɔk/. Bangla has the same type of tion needs to be converted into ASCII to incorpo-
non-natural language ambiguity like Hindi [K. rate into festival.
Panchapagesan et al. 2004] in the token year- We have implemented the G2P algorithm that
number and time-floating number. For example: is proposed by Ayesha et al. [Ayesha et al. 2006]
(i). the token ১৯৯৮ (1998) could be considered to handle unknown words and proper name. In
as a year and at the same time it could be consi- Festival, the UTF-8 textual input was converted
dered as number and (ii). the token ১২. ৮০ into ASCII based phonetic representation in a
(12.80) could be considered as a floating point Festival’s context sensitive rule [A.W. Black et
number and it could be considered as a time. al. 2003]. The rules were re-written in UTF-8

157
multi-byte format following the work done for more likely between content words and function
Telugu [C. Kamisetta et al. 2006] and Sinhala words. A rule is defined to predict a break if the
[Ruvan et al. 2007]. The method was proven to current word is a content word and the next is
work well with promising speed. The rules pro- seemingly a function word and the current word
posed in [Ayesha et al. 2006] were expanded up is more than 5 words from a punctuation symbol.
to 3880 rules when re-written in Festival context Since function words are limited in a language so
sensitive format. we specified them as function words and consi-
Another attempt has been made to reduce the dered rest of them as content words. The func-
size of the lexicon for TTS. The lossless com- tion words that we used here to implement the
pression [Paul 2009] technique was applied to phrase break model is shown in Table 4. These
reduce the size of the lexicon. Lossless compres- function words need to be converted into ASCII
sion technique is a technique where the output is form to incorporate into festival phrase breaking
exactly the same as when the full lexicon is used. algorithm.
It is just the generalities of the lexicon that has
been exactly captured in a set of rules. This tech- Function words
nique reduces the size of our lexicon to ~50K অ, অতএফ, অথচ, অথফা, অমধকন্তু, অল঩ক্ষা,
entries from 93K. অথথাৎ, আয, আযও, এ, এই, এফ , ও, মক ফা,
3.3.4 Syllabification and stress মকন্তু, তথা, তথাম঩, তফু, তফুও, তাই, মতা,
নইলর, নইলর, নতুফা, ন৞লতা, না- ঴৞, ফলট, ফয ,
Festival’s default syllabification algorithm based ফযঞ্চ্য, ফস্তুত, ফা, মথা, মমদ, মমদও, মম, মমন,
on sonority sequencing principle [A.W. Black et মমল঴তু, সুতযা , ঴ঠাৎ, ঴৞লতা
al. 2003] is used to syllabify the Bangla words.
Besides the default syllabification algorithm, our Table 4: Function words
lexicon has also been syllabified along with pro-
nunciation. To predict accent and boundary tone, Festival
Little work has been done on Bangla stress. Iden- uses simple rules to produce sophisticated sys-
tifying the stress pattern for Bangla is beyond the tem. To make a more sophisticated system such
scope of this paper. Considering Bangla as a as a statistical model one needs to have an ap-
stress less language we have used Festival’s de- propriate set of data. Due to the lack of the avail-
fault stress algorithm. In our implementation of ability of this data we used a simple accent pre-
lexicon we have not incorporated the stress diction approach [A.W. Black et al. 2003] which
marker. proved surprisingly well for English. This ap-
3.3.5. Prosody Implementation proach assigns an accent on lexically stressed
syllable in all content words.
Prosody is one of the important factors contribut- Festival uses different approach for F0 genera-
ing to natural sounding speech. This includes tion such as F0 by rule, CART tree and tilt mod-
phrasing, accent/boundary prediction, duration eling. In our implementation we used rule based
assignment to phones and f0 generation. The approach. An attempt has been made to make a
presence of phrase breaks in the proper positions CART tree based model from the data; however,
of an utterance affects the meaning, naturalness surprisingly that has not been work well.
and intelligibility of the speech. Festival supports Several duration models support by Festival such
two methods for predicting phrase breaks. The as fixed models, simple rules models, complex
first one is to define a Classification and Regres- rules models and trained models. We used fixed
sion Tree (CART). The second and more elabo- duration model that was implemented from the
rate method of phrase break prediction is to im- work done by Firoj et al. [Firoj et al. 2008
plement a probabilistic model using probabilities (b)][Firoj et al. 2008 (a)].
of a break after a word, based on the part of
speech of the neighboring words and the pre- 3.4 Development of diphone database
vious word [A.W. Black et al. 2003]. However, Developing a speech database is always time
due to the lack of a POS tagger for Bangla, we consuming and laborious. The basic idea of
have not able to construct a probabilistic model building a diphone database is to explicitly list
yet. Therefore, we decided to use a simple CART all phone-phone combination of a language. It is
based phrase breaking algorithm described in mentioned in section 3.2 that Bangla language
[A.W. Black et al. 2003]. The algorithm is based has 30 consonants and 35 vowels (monophthong,
on the assumption that phrase boundaries are

158
diphthong) phonemes. In general, the number of The most laborious and painstaking task is to
diphone in a language is the square of the num- clean the recording and then hand-labeled the
ber of phones. Since Bangla language consists of diphone using the speech analysis software tool
65 phones, so the number of diphones are ‘Praat’ 4 . During labeling, at first we labeled
(65X65) 4225. In addition, silence to phones are phone boundary, then automatically marked the
(1X65) 65, phones to silence are (65X1) 65 and a diphone boundary using Praat script. Another
silence. So the total number of diphones is 4336. important factor is that, every boundary should
In the first step, a list has been made to maintain be placed in zero crossing. Failing to do so pro-
all the possible vowel consonant combination duces audible distortions, this in turns generates
with the following pattern: VC, CV, VV, CC, clicks. Afterwards, a script was written to trans-
SIL_V, SIL_C, V_SIL, C_SIL and SIL. Here form Praat textgrid files into diphone index file
SIL is silence, V is vowel and C is consonant. (.est) [A.W. Black et al. 2003] as required by
Silence is considered as a phoneme, usually tak- Festival.
en at the beginning and ending of the phonemes Festival, in its publicly distributed form only
to match the silences occurring before, between supports residual excited Linear Predictive Cod-
and after the words. They are therefore an im- ing (LPC). This method requires pitch marks,
portant unit within the diphone inventory. These LPC parameters and LPC residual values for
diphones were embedded with carrier sentences each diphone in the diphone database. The script
using an external program. The diphone is in- make_pm_wave provided by speech tools [A.W.
serted in the middle of the word of a sentence, Black et al. 2003] was used to extract pitch
minimizing the articulatory effects at the start marks from the wave files. Then, the make_lpc
and end of the word. Also, the use of nonsense command was invoked in order to compute LPC
words helped the speaker to maintain a neutral coefficients and residuals from the wave files
prosodic context. Though there have been vari- [A.W. Black et al. 2003]. To maintain an equal
ous techniques to embed diphone with carrier power we used proprietary software tool to nor-
sentences, here nonsense words were used to malize it in terms of power so that all diphones
form carrier sentences [A.W. Black et al. 2003]. had an approximately equivalent power. After
In this list, there could be redundant diphones that the diphone database was grouped in order
those need to be marked and omitted. The study to make it accessible by Festival’s UniSyn syn-
of phonotactics says that all phone-phone pair thesizer module, and to make it ready for distri-
cannot be exist in a language. Due to the lack of bution.
existing work and linguistic experts we were not
able to work on this phenomenon. Therefore, the 4 Integration with applications
whole diphone list was selected for recording.
The Bangla Text to Speech runs on Linux,
Since speaker choice is perhaps one of the
Windows and Mac OSX. There is also a web-
most vital areas for recording so a careful meas-
enabled front-end for the TTS, making this tool
ure had taken. Two potential speakers was cho-
available at anytime and from anywhere.
sen and their recording were played to a listening
Since Festival is incapable of reading UTF-8
group and asked them which they prefer. Ac-
text files with byte-order marker (BOM) so ma-
cording to the measurement of the listening
nual BOM removal patch was used which was
group a male speaker was chosen who is a pro-
written by Weerasinghe et al. [Ruvan et al. 2007].
fessional speaker and aged 29.
This patch was incorporated with Festival text
As far as recording conditions is concerned, we
processing module.
tried to maintain as high quality as possible. The
To develop windows version we had moti-
speech data was digitized at a sample rate 44.1
vated by the work carried out in the Welsh and
kHz, sample width 24-bit resolution and stored
Irish Speech Processing Resources (WISPR)
as wave format. After each recording, the mod-
project [B. Williams et al. 2006]. Following the
erator checked for any misleading pronunciation
work of WISPR, we implemented TTS using
during the recording, and if so, the affected utter-
Microsoft Speech Application Programming In-
ances were re-recorded.
terface (MS-SAPI) which provides the standard
There were a few challenges in the recording.
speech synthesis and speech recognition inter-
First, speaker was asked to keep the speaking
face within Windows applications [Microsoft
style consistent. Second, speaker was supervised
to keep the same tone in the recording.
4
Available from: http://www.praat.org

159
1999]. Consequently, the MS-SAPI compliant intelligibility test, we have performed a unit test
Bangla voice is accessible via any speech on text normalizer and G2P converter. The per-
enabled Windows application. The system has formance of text normalizer is 87% only for am-
been tested with NVDA 5 and Dolphin 6 screen biguous tokens and that of G2P converter is
reader. Moreover, it is also tested with Word- 89%.
Talk7, a free text-to-speech plug-in for Microsoft
Word which runs as a macro. Currently Bengali 6 Conclusions
speaking print disabled community accessing
Here the development of the first-ever complete
local language content using Bangla Text to
Text to Speech (TTS) system has described, that
speech system via screen reader.
can convert a Unicode encoded Bangla text into
Besides, there are few other applications that
human speech. It is distributed under an open
currently testing this system such as talking dic-
source license to empower both the user and de-
tionary, DAISY8 book, agro-information system
veloper communities. This TTS system can also
and news reader. Using this system one of the
be used with any available Screen Reader. In
newspapers in Bangladesh developed their audio
addition to the standalone TTS client, it can be
version of newspaper to make mp3 of their daily
integrated into virtually any application, and can
content.
also be accessed as a Web Service. Incorporating
5 Evaluation this technology in various applications such as
screen reader for the visually impaired, touch
Any real system needs to undergo rigorous test- screen based agro-information system, talking
ing before deployment. Though TTS testing is books, telecenter applications, e-content, etc.,
not a simple or widely agreed area, it is widely can potentially bridge the literacy divide in Ban-
agreed that a TTS system has two main goals on gladesh, which in turn goes towards bridging the
system test; that is a synthesized speech should digital divide. An evaluation of the system has
be i) intelligible and ii) natural. Intelligibility test been done based on MRT and unit testing on a
can be performed by word recognition tests or few components to check intelligibility.
comprehension tests where listeners are played a Since the voice developed here is diphone conca-
few words either in isolation or in a sentence and tenation based and it lacks proper intonation
asked which word(s) they heard. In naturalness modeling so it produces robotic speech. There-
test, listeners are played some speech (phrase or fore, a natural sounding voice needs to be made
sentence) and simply asked to rate what they in future, which could be performed by develop-
hear. This can be done by mean opinion score. ing a unit selection voice. Besides that, a few
Since these testing may not always be the best works need to be done in future to improve the
approach so people also use unit testing ap- intelligibility of the system such as POS tagger,
proach. improvement of G2P algorithm, improvement of
As our goal was to make a general-purpose syn- text normalizer and working on intonation mod-
thesizer, a decision was made to evaluate it under eling.
the intelligibility criterion and unit testing on a Acknowledgments
few components. The most commonly used word This work has been supported in part by the PAN
recognition test - modified rhyme test (MRT) Localization Project (www.panl10n.net), grant
[Paul 2009] was designed to test Bangla TTS from the International Development Research
system. Based on the MRT we designed a set of Center (IDRC), Ottawa, Canada. We would also
77 groups - 5 words each. Therefore a set of 385 like to thank Dr Sarmad Hussain (NUCES), and
words came into testing. The words in each Naira Khan (Dhaka University).
group are similar and differ in only one conso-
nant and the users were asked to account which References
word they have heard on a multiple choice sheet.
A.G. Ramakishnan, K. Bali, P.P. Talukdar and N.S.
Based on the test the overall intelligibility of the
Krishna, 2004, Tools for the Development of a
system from 6 listeners is 96.96%. Besides the HindiSpeech Synthesis System, In 5th ISCA
Speech Synthesis Workshop, Pittsburgh, 2004, pp.
5 109-114.
NVDA – NonVisual Desktop Access. www.nvda-
project.org/ A.W. Black, and K.A. Lenzo, 2003, Building Synthet-
6
Dolphin screen reader.www.yourdolphin.com/ ic Voices, Language Technologies Institute, Car-
7
Wordtalk - www.wordtalk.org.uk/ negie Mellon University and Cepstral LLC. Re-
8
DAISY - Digital Accessible Information System trieved from: http://festvox.org/bsv/

160
Aniruddha Sen, 2004, Bangla Pronunciation Rules Bilinear Trans-forms", IEEE Trans. Speech and
and a Text-to-Speech System, Symposium on In- Audio Proc., 7(6):697-708.
dian Morphology, Phonology & Language Engi- Ö. Salor, B. Pellom and M. Demirekler, 2003, Im-
neering, pp. 39. plementation and Evaluation of a Text-to-Speech
Asok Bandyopadhyay, 2002, Some Important Aspects Synthesis System for Turkish, Proceedings of Eu-
of Bengali Speech Synthesis System IEMCT Pune, rospeech-Interspeech, Geneva, Switzerland, pp.
June 24-25 . 1573-1576.
Ayesha Binte Mosaddeque, Naushad UzZaman and Paul Taylor, 2009, Text-to-Speech Synthesis, Univer-
Mumit Khan, 2006, Rule based Automated Pro- sity of Cambridge, February.
nunciation Generator, Proc. of 9th International Prothom-Alo, 2009, www.prothom-alo.com, Daily
Conference on Computer and Information Tech- Bengali newspaper
nology (ICCIT 2006), Dhaka, Bangladesh, Decem- Python, 2008, www.python.org, version 2.5.2
ber 2006. R.J. Jones, A. Choy and B. Williams, 2006, Integrat-
B. Williams, R.J. Jones and I. Uemlianin, 2006, Tools ing Festival and Windows, InterSpeech 2006, 9th
and Resources for Speech Synthesis Arising from a International Conference on Spoken Language
Welsh TTS Project, Fifth Language Resources and Processing, Pittsburgh, USA.
Evaluation Conference (LREC), Genoa, Italy, Ruvan Weerasinghe, Asanka Wasala, Viraj Welgama
2006. and Kumudu Gamage, 2007, Festival-si: A Sinhala
C. Kamisetty and S.M. Adapa, 2006, Telugu Festival Text-to-Speech System, Proceedings of Text,
Text-to-Speech System, Retrieved from: Speech and Dialogue, 10th International Confe-
http://festivalte.sourceforge.net/wiki/Main_Page rence, TSD 2007, Pilsen, Czech Republic, Septem-
CRBLP, 2010, CRBLP pronunciation lexicon, [On- ber 3-7, Page 472-479.
line], Available: http://crblp.bracu.ac.bd/demo/PL/ S.P. Kishore, R. Sangal and M. Srinivas, 2002, Build-
Firoj Alam, Promila Kanti Nath and Mumit Khan, ing Hindi and Telugu Voices using Festvox, Pro-
2007, Text To Speech for Bangla Language using ceedings of the International Conference On Natut-
Festival, Proc. of 1st International Conference on al Language Processing 2002 (ICON-2002), Mum-
Digital Communications and Computer Applica- bai, India.
tions (DCCA2007), Irbid, Jordan Shyamal Kr. DasMandal, Barnali Pal , 2002, Bengali
Firoj Alam, S. M. Murtoza Habib and Mumit Khan, text to speech synthesis system a novel approach
2008 (a), Research Report on Acoustic Analysis of for crossing literacy barrier, CSI-YITPA(E)
Bangla Vowel Inventory, Center for Research on Tanuja Sarkar, Venkatesh Keri,. Santhosh M and Ki-
Bangla Language Processing, BRAC University. shore Prahallad, 2005, Building Bengali Voice Us-
Firoj Alam, S. M. Murtoza Habib and Mumit Khan, ing Festvox, ICLSI.
2009, Text Normalization System for Bangla, Con- Thierry Dutoit, 1997, An Introduction to Text-To-
ference on Language and Technology 2009 Speech Synthesis, Kluwer Academic Publishers
(CLT09), NUCES, Lahore, Pakistan, January 22- Wikipedia contributors. Bengali script [Internet]. Wi-
24. kipedia, 2010, The Free Encyclopedia; 2010 Jul 22,
Firoj Alam, S. M. Murtoza Habib and Professor Mu- 17:27 UTC [cited 2010 Jul 30], Available online
mit Khan, 2008 (b), Acoustic Analysis of Bangla at:
Consonants, Proc. Spoken Language Technologies http://en.wikipedia.org/w/index.php?title=Bengali_
for Under-resourced language (SLTU’08), Viet- script&oldid=374884413.
nam, May 5-7, page 108-113. www.somewhereinblog.net, 2008, Blog
forum.amaderprojukti.com, 2008, Forum
K. Panchapagesan, Partha Pratim Talukdar, N. Srid-
har Krishna, Kalika Bali, A.G. Ramakrishnan,
2004, Hindi Text Normalization, Fifth Internation-
al Conference on Knowledge Based Computer Sys-
tems (KBCS), Hyderabad, India, 19-22 December
2004. Retrieved (June, 1, 2008).
Khair Md. Yeasir Arafat Majumder, Md. Zahurul
Islam, Naushad UzZaman and Mumit Khan, 2006,
Analysis of and Observations from a Bangla News
Corpus, in proc. 9th International Conference on
Computer and Information Technology, Dhaka,
Bangladesh, December.
Microsoft Corporation.: Microsoft Speech SDK Ver-
sion 5.1, 1999, Retrieved from:
http://msdn2.microsoft.com/ens/library/ms990097.
aspx. Smith, J. O. and Abel, J. S., Bark and ERB

161