Text Processing for Text-to-SpeechSystems in Indian Languages

Anand Arokia Raj 1 , Tanuja Sarkar 1 , Satish Chandra Pammi 1 ,

Santhosh Yuvaraj 1 , Mohit Bansal 2 , Kishore Prahallad1 3 , Alan W Black3

InternationalInstitute of Information Technology, Hyderabad, India.

1
2
Indian Institute of Technology, Kanpur, India.
3
Language Technologies Institute, Carnegie Mellon University, Pittsburgh, USA.
skishore@cs.cmu.edu, awb@cs.cmu.edu

Abstract linguistic resources), it involves several complexities starting

from accumulation of text corpora in digital and processable
To build a natural sounding speech synthesis system, it is essen- format. Linguistic components are not available in such rich
tial that the text processing component produce an appropriate fashion for all languages of the world. In practical world, mi-
sequence of phonemic units correspondingto an arbitrary input nority languages including some of the Indian languages do not
text. In this paper we discuss our efforts in addressingthe issues have that luxury of assuming some or any of the linguisticcom-
of Font-to-Aksharamapping, pronunciationrules for Aksharas, ponents.
text normalizationin the context of building text-to-speechsys- The purpose of this paper is to describe our efforts at IIIT
tems in Indian languages. Hyderabad to build a generic framework for build text process-
ing modules and linguistic resources which could be extended
1. Introduction to all of the Indian languages with minimal efforts and time.
Our approach is to make use of minimal language informa-
The objective of a text to speech system is to convert an arbitrary
tion (i.e., information available with an average educated native
given text into a correspondingspoken waveform. Text process-
speakers), take the aid of acoustic data and machine learning
ing and speech generation are two main components of a text to
techniques [5]. In this paper we summarize some of our ef-
speech system. The objective of the text processing component
forts in this direction but mainly for font identification,Font-to-
is to process the given input text and produce appropriate se-
Akshara conversion, pronunciation rules for Aksharas and text
quence of phonemic units. These phonemic units are realized
normalization.
by the speech generation component either by synthesis from
parameters or by selection of a unit from a large speech corpus.
For natural sounding speech synthesis, it is essential that the 2. Nature of Indian Language Scripts
text processing component produce an appropriate sequence of The scripts in Indian languageshave originatedfrom the ancient
phonemic units correspondingto an arbitrary input text. Brahmi script. The basic units of the writing system are referred
One of the question often asked by end-users is why we to as Aksharas. The properties of Aksharas are as follows: (1)
don’t have TTS systems for all or many of the 23 official Indian An Akshara is an orthographicrepresentationof a speech sound
languages. What are the complexities: Is it because the syn- in an Indian language; (2) Aksharas are syllabic in nature; (3)
thesis technology isn’t matured enough to be able to build for The typical forms of Akshara are V, CV, CCV and CCCV, thus
any language or is it because of the non-existence of speech have a generalized form of C*V.
databases in Indian languages?. Unfortunately, for a decade The shape of an Akshara depends on its composition of
the core speech generationtechnologyi.e., generationof speech consonants and the vowel, and sequence of the consonants. In
from a phonemic sequence has largely been automated due to defining the shape of an Akshara, one of the consonant symbols
unit selection techniques [1]. With the introduction of statisti- acts as pivotal symbol (referred to as semi-full form). Depend-
cal parametric speech synthesis techniques, it is much easier to ing on the context, an Akshara can have a complex shape with
build a voice in a language with fewer sentences and a smaller other consonant and vowel symbols being placed on top, be-
speech corpus [2] [3]. low, before, after or sometimes surrounding the pivotal symbol
It is difficult to convince an end-user that the input to a (referred to as half-form).
TTS system is not a phonemic sequence but rather the raw Thus to render an Akshara, a set of semi-full or half-forms
text as available in news websites, blogs, documents etc which have to be rendered, which in turn are rendered using a set of
contain the required text in font-encodings, native scripts and basic shapes referred to as glyphs. Often a semi-full form or
non-standard words such as addresses, numbers, currency etc. half-form is rendered using two or more glyphs, thus there is no
The majority of the issues are associated in building a TTS one-to-one correspondencebetween glyphs of a font and semi-
for a new language is associated with handling of real-world full or half-forms [6].
text [4]. Current state-of-art TTS system in English and other
well-researched languages use such rich set of linguistic re-
2.1. Convergence and Divergence
sources such as word-sense disambiguation,morphological an-
alyzer, Part-of-Speech tagging, letter-to-sound rules, syllabifi- There are 23 official languages of India, and all of them ex-
cation, stress-patterns in one form or the other to build a text cept English and Urdu share a common phonetic base, i.e., they
processing component of a TTS system. However for minority share a common set of speech sounds. While all of these lan-
languages(which are not well researchedor do not have enough guages share a common phonetic base, some of the languages

6th ISCA Workshop on Speech Synthesis, Bonn, Germany, August 22-24, 2007 188
such as Hindi, Marathi and Nepali also share a common script type. In this approach, the term refers to a ’glyph’ and the docu-
known as Devanagari. But languages such as Telugu, Kannada ment refers to the font-data of a particular ’font-type’. Here the
and Tamil have their own scripts. glyph-sequencecould mean a single glyph or ’current and next’
The property that makes these languages separate can be glyph or ’previous, current and next’ glyph etc.
attributed to the phonotactics in each of these languages rather To build a document for each font-type, a web-site for
than the scripts and speech sounds. phonotactics is the permis- each font-type was manually identified and around 0.12 million
sible combinationsof phones that can co-occur in a language. unique words were crawled for each of the font-type. The set of
unique words for each font-type are referred to as a document
2.2. Digital Storage of Indian Language Scripts representing the particular font-type. Thus given N documents
(each representing a font-type), we considered three different
There is a chaos as far as the text in Indian languages in elec- terms namely, a single glyph or current and next glyph or pre-
tronic form is concerned. Neither can one exchange the notes vious, current and next glyph. For each term a TF-IDF weight
in Indian languages as conveniently as in English language, nor was obtained as follows: (i) Calculate the term frequency for
can one perform search easily on texts in Indian languagesavail- the glyph-sequence: The number of times that glyph-sequence
able over the web. This is because the texts are being stored in occurred divided by the total number of glyph-sequences in
ASCII font dependent glyph codes as opposed to Unicode. that specific document. (ii) Calculate document frequency: In
The glyph coding schemes are typically different for differ- how many different documents (font-types)that specific glyph-
ent languages and within a language there could exists several sequence has occurred. (iii) Calculate inverse document fre-
font-types with their own glyph codes (as many as major news- quency of the term and take logarithm of inverse document fre-
portals in a language). To view the websites hosting the content quency.
in a particular font-type,these fonts have to be installedon local To identify the font-type of a given test font-data, the
machine. As this was the technology existed before the era of steps involved are as follows: 1) Generate the terms (glyph-
Unicode and hence a lot of electronic data in Indian languages sequences)of the test font-data2) Compute the relevancy scores
were made and available in that form [7]. of the terms and for each of the document (font-type) using
the corresponding TF-IDF weights of the terms 3) The test
2.3. Need for Handling Font-Data font-data belongs to the document (font-type)which produces a
The text available in a font-encoding (or font-type) is referred maximum relevancy score.
to as font-data. While Unicode based news-portals and web- The performance of TF-IDF approach for identification of
pages are increasing, there are two main reasons to deal with font-type was evaluated on 1000 unique sentences and words
ASCII based font-data: 1) Given that there are 23 official In- per font-type. We have added English data as also one of
dian languages,and the amount of data available in ASCII based the testing set, and is referred to as English-text. The perfor-
font-encodingsis much larger than the text content available in mance of font-type identification system using different terms
Unicode format, 2) If a TTS system has to read the text from single glyph, current and next glyphs, previous, current and
a ASCII font based website then the TTS system should au- next glyphs are shown in Table 1, Table 2 and Table 3 respec-
tomatically identify the font-type and process the font-data to tively and it could be observed that the use of previous, current
generate speech. and next glyphs as a term provided an accuracy of 100% in iden-
tification of font-type even at the word level.
2.4. A Phonetic Transliteration Scheme for Digital storage
of Indian Language Scripts
Table 1: Performance of Single glyph based font models
To handle diversified storage formats of scripts of Indian lan- Font Name Sentence-Level Word-Level
guages such as ASCII based fonts, ISCII (Indian Standard code Amarujala (Hindi) 100% 100%
for Information Interchange) and Unicode etc, it is useful and Jagran (Hindi) 100% 100%
becomes necessary to use a meta-storageformat.
Webdunia (Hindi) 100% 0.1%
A transliterationscheme maps the Aksharas of Indian lan- SHREE-TEL (Telugu) 100% 7.3%
guages onto English alphabets and it could serve as meta-
Eenadu (Telugu) 0% 0.2%
storage format for text-data. Since Aksharas in Indian lan-
Vaarttha (Telugu) 100% 29.1%
guages are orthographic represent of speech sound, and they
Elango Panchali (Tamil) 100% 93%
have a common phonetic base, it is suggested to have a pho-
netic transliterationscheme such as IT3 [8] [6]. Thus when the Amudham (Tamil) 100% 100%
font-data is converted into IT3, it essentially turns the whole SHREE-TAM (Tamil) 100% 3.7%
effort into font-to-Aksharaconversion. English-text 0% 0%

3. Identificationof Font-Type
Given a document we often need to identify the font-type, and
4. Font-to-AksharaMapping
sometimes a document can contain the data encoded in differ- Font-data conversion can be defined as converting the font en-
ent font-types. Then the task would boil down to identifyingthe coded data into Aksharas representedusing phonetic translitera-
font-type for each line or for each word. In this paper, we pro- tion scheme such as IT3. As we already mentioned that Aksha-
pose the use of TF-IDF approach for identificationof font-type. ras are split into glyphs of a font, and hence a conversion from
The term frequency - inverse document frequency (TF-IDF) ap- font-data has essentially to deal with glyphs and model how a
proach is used to weigh each glyph-sequence in the font-data sequence of glyphs are merged to form an Akshara. As there ex-
according to how unique it is. In other words, the TF-IDF ap- ist many fonts in Indian languages, we have designed a generic
proach captures the relevancy among glyph-sequenceand font- framework has been designed for the conversion of font-data. It

6th ISCA Workshop on Speech Synthesis, Bonn, Germany, August 22-24, 2007 189
 (viii) Schwa Deletion.
Table 2: Performanceof current and next glyph based font mod-
The Modifier Modificationis the process where the charac-
els
Font Name Sentence-Level Word-Level ters get modified because of the language modifiers like virama
and nukta (ka + virama = k1). The Language Preprocessing
Amarujala (Hindi) 100% 100%
step deals with some language specific processing like (aa3 +
Jagran (Hindi) 100% 100% i3 = ri in Tamil) and (r4 moves in front of the previous first full
Webdunia (Hindi) 100% 100% consonant in Hindi). The Consonant Assimilation is known as
SHREE-TEL (Telugu) 100% 100% getting merged two or more consonant glyphs and forms a valid
Eenadu (Telugu) 100% 100% single consonant like (d1 + h5 = dh1 in Telugu). The Maatra
Vaarttha (Telugu) 100% 100% Assimilation is known as getting merged two or more maatra
Elango Panchali (Tamil) 100% 100% glyphs and forms a valid single maatra like (aa3 + e4 = o3 in
Amudham (Tamil) 100% 100% Hindi). The Consonant-Vowel Assimilationis known as getting
SHREE-TAM (Tamil) 100% 100% merged two or more consonant and vowel glyphs and forms a
English-text 100% 96.3% valid single consonant like (e + a4 + u5 = pu in Telugu). The
Vowel-Maatra Assimilation is known as getting merged two or
more vowel and maatra glyphs and forms a valid single vowel
Table 3: Performance of previous, current and next based font like (a + aa3 = aa in Hindi). The ConsonantClusteringin known
models as merging the half consonant which usually occurs at the bot-
Font Name Sentence-Level Word-Level tom of a full consonant to that full consonant like (la + l5 = lla
Amarujala (Hindi) 100% 100% in Hindi). The Schwa Deletion is deleting the inherent vowel
Jagran (Hindi) 100% 100% ’a’ from a full consonant in necessary places like (ka + ii3 =
Webdunia (Hindi) 100% 100% kii).
SHREE-TEL (Telugu) 100% 100%
Eenadu (Telugu) 100% 100% 4.3. Testing and Evaluation
Vaarttha (Telugu) 100% 100% The evaluation on these font converters is carried out in two
Elango Panchali (Tamil) 100% 100% phases. We picked up three different font-types for training or
Amudham (Tamil) 100% 100% forming the assimilationrules and one new font-type for testing
SHREE-TAM (Tamil) 100% 100% per language. In the first phase for the selected three font-types
English-text 100% 100% the assimilation rules are formed and refined. In the second
phase we chose a new font-type and built the base-map table
only and used the existing converter without any modifications.
has two phases, in the first phase we are building the base-map We have taken 500 unique words per font-type and generated
table for a given font-type and in the second phase forming and the conversion output. The evaluation results in Table 4 show
ordering the assimilationrules for a specific language. that the font converter performs consistently even for a new
font-type. So it is only sufficient to provide the base-map table
4.1. Building a Base-Map Table for a Font-type for a new font-type to get a good conversion results. The issue
of Font-to-Akshara mapping has been attempted in [7] and [9]
The base-map table provides the mapping basic between the but we believe that our framework is a generic one which could
glyphs of the font-type to the Aksharas represented in IT3 easily be extended to a new font-type with > 99% conversion
transliteration scheme. The novelty in our mapping was that accuracy.
the shape of a glyph was also included in building this mapping
table. The shape of a glyph is dictated by whether it is rendered
as pivotal consonant, or on top, bottom, left or right of the piv- 5. Building Pronunciation Models For
otal consonant. Thus the pivotal glyphs were appended with 0 Aksharas
(for full characters such as e, ka) or 1 (for half consonants such
as k1, p1), ’2’ for glyphs occur at left hand side of a basic char- Having converted the font-data into Aksharas, the next step is
acter (ex: i2, r2), ’3’ for glyphs occur at right hand side of a to obtain appropriate pronunciation for each of the Aksharas.
basic character (ex: au3, y3), ’4’ for glyphs occur at top of a As noted earlier, Aksharas are orthographic representation of
basic character (ex: ai4, r4) and ’5’ for glyphs occur at bottom speech sounds and it is commonly believed or quoted that there
of a basic character (ex: u5, t5). is direct correspondence between what is written and what is
spoken in Indian languages, however, there is no one-to-one
correspondence between what is written and what is spoken.
4.2. Forming AssimilationRules
Often some of the sounds are deleted such as Schwa deletion
In the conversion process the above explained basic-mapping in Hindi. Schwa is the default short vowel /a/which is associ-
table will be used as the seed. A well defined and ordered ated with a consonant, and often it is deleted to aid in faster
set of assimilation rules have to be formed for each and every pronunciation of a word. Similarly there exists exceptions for
language. Assimilation is the process of merging two or more Bengali and Tamil. There are attempts to model these excep-
glyphs and generating a valid single character. This assimilation tions in the form of the rules, however, they are often met with
happens at different levels and our observation across many lan- limited success or they use linguistic resources such as Morph
guages was that the firing of following assimilation rules were analyzer. Such linguistic resources may not always be available
universally applicable. The rules are:(i) Modifier Modification, for minority languages. Thus we had built a framework based
(ii) Language Preprocessing, (iii) Consonant Assimilation, (iv) on machine learning techniques where pronunciationof Aksha-
Maatra Assimilation, (v) Consonant-Vowel Assimilation, (vi) ras could be modeled using machine learning techniques and
Vowel-Maatra Assimilation, (vii) Consonants Clustering and using a small set of supervised training data.

6th ISCA Workshop on Speech Synthesis, Bonn, Germany, August 22-24, 2007 190
 to predict the pronunciation of a phone. Table 5 shows the re-
Table 4: Performance results for font conversion in Indian lan- sults of pronunciationmodel for Hindi, Bengali and Tamil using
guages various level of contextual features. We found that that a con-
Language Font Name Training/Testing Accuracy text level of 4 (i.e., 4 phones to the left and 4 phones to the right)
Hindi Amarujala Training 99.2% was sufficient to model the pronunciation and moving beyond
Jagran Training 99.4% the level of 4, the performance was degraded.
Naidunia Training 98.8%
Webdunia Training 99.4%
Chanakya Testing 99.8% Table 5: Pronunciation Model with Contextual features
Languages Context Level
Marathi Shree Pudhari Training 100% 2 3 4 6
Shree Dev Training 99.8% Hindi 90.24% 91.44% 91.78% 91.61%
TTYogesh Training 99.6% Bengali 82.77% 84.48% 84.56% 83.56%
Shusha Testing 99.6% Tamil 98.16% 98.24% 98.10% 98.05%

Telugu Eenadu Training 93%

Vaartha Training 92% 5.3. Acoustic-Phoneticand Syllabic Features
Hemalatha Training 93%
TeluguFont Testing 94% Acoustic phonetic features lists the articulatorypropertiesof the
consonants and the vowels. Typically vowels are characterized
Tamil Elango Valluvan Training 100% by the front, back, mid position of the tongue while consonants
Shree Tam Training 99.6% are characterizedby manner and place of articulation and voic-
Elango Panchali Training 99.8% ing and nasalization features. Syllabic features indicate where
a particular syllable is of type CV or CCV, or CVC etc. The
Tboomis Testing 100%
performance of the pronunciation model for Hindi, Tamil and
Bengali using syllabicand acoustic-phoneticfeaturesof the cur-
Kannada Shree Kan Training 99.8%
rent and neighboring phones are shown in Table 6 and Table
TTNandi Training 99.4% 7 respectively. We found that the use of syllabic or acoustic-
BRH Kannada Training 99.6% phonetic features didn’t show any significant improvement than
BRH Vijay Testing 99.6% that of contextual features for Hindi, Tamil and Bengali.
A rule based algorithm for Hindi LTS is given in [10]. To
Malayalam Revathi Training 100% compare our results with the rule-basedalgorithm,we have used
Karthika Training 99.4% the same algorithm with out morphologicalanalyzer on our test
Thoolika Training 99.8% data set. We found that the performanceof pronunciationmodel
ShreeMal Testing 99.6% using rule-based technique was 88.17%. while the decision for-
est model in Table 6 was providing an accuracy of 92.29%.
Gujarati Krishna Training 99.6%
Krishnaweb Training 99.4%
Gopika Training 99.2% Table 6: Pronunciation Model with Syllabic features
Divya Testing 99.4% Feature Sets Languages
Hindi Bengali Tamil
Syl Struct. of Cur. Phone 92.29% 82.41% 98.31%
5.1. Creation of Data-set Syl Struct. of all Phones 91.61% 67.56% 98.27%

Given the input word list with the correspondingpronunciations

in terms of phones, feature vectors were extracted for training
the pronunciation model at the phone level. About 12200 sen- Table 7: Pronunciation Model with Acoustic-Phoneticfeatures
tences in IT3 format were used to collect the training data, for
Feature Sets Languages
building the pronunciationmodel in Hindi. These sentenceshad
about 26000 unique words, which were used to extract around Hindi Bengali Tamil
32800 feature vectors. Different sets of feature vectors to ex- Acoustic Phonetic 89.73% 84.78% 98.18%
periment on the selection of features. As for Bengali and Tamil, + Syl Struct. of Curr. Phone 89.73% 81.21% 98.17%
5000 words with corresponding pronunciations were used for + Syl Struct. of all Phones 91.09% 69.33% 98.13%
obtaining about 9000 feature vectors.

5.2. Use of Contextual Features 6. Normalizing of Non-StandardWords

Contextual features refers to the neighbor phones in a definite Unrestrictedtexts include Standard Words (common words and
window-size/level. Using the contextual features, experiments Proper Names) and Non-Standard Words (NSWs). Standard
were performed for various Contextual Levels (CL). A decision Words have a specific pronunciation that can be phonetically
forest was built for each phone to model its pronunciation. A described either in a lexicon, using a disambiguation process-
decision forest is a set of decision trees built using overlapping ing to some extent, or by letter-to-sound rules. In the context of
but different sub-sets of the training data and it employs a ma- TTS the problem is to decide how an automatic system should
jority voting scheme on individual prediction of different trees pronounce a token; even before the pronunciationof a token, it

6th ISCA Workshop on Speech Synthesis, Bonn, Germany, August 22-24, 2007 191
 Table 8: Taxonomy of NSWs with examples Table 9: Performance of prediction of NSW-category Using
Category Description Examples Word Level Features
Addr Address 12/451 Language % accuracy on % accuracy on
(house/streetno.) Janapath Road Training set TS1
Curr Currency Rs. 7635.42 Telugu 99.57% 63.52%
Count Count of items 10 computers, Hindi 99.80% 66.99%
500 people Tamil 99.01% 55.42%
Date Date(to be expanded) 1/1/05, 1997-99
PhoneNo As sequence of digits 040 2300675
Pin As sequence of digits 208023 Table 10: Performance of prediction of NSW-category Using
Score Cricket, tennis scores India 123/4, Syllable level Features
sets 3-5 3-4 5-6 Language % accuracy on % accuracy on Diff with
Time Time (to be expanded) 1.30, 10:45-12:30, Training set TS1 base-line
11.12.05, 1930 hrs Telugu 99.57% 91.00% 27.48%
Units As decimal or number 10.5 kms, 98 %, Hindi 99.80% 82.80% 15.81%
13.67 acres Tamil 99.01% 87.20% 31.78%
NUM Default category

(TS1). The results of prediction of NSW category on TS1 is

shown in Table 9.
is important to identify the NSW-Category of a token. A typical
The performance of the base-line system on TS1 is around
set of NSW-category and their examples are shown in Table 8.
60%. After analyzing the errors made by the system, we found
that the errors are primarily due to new words found in the con-
6.1. Creation of Supervised Training Data
text of NSW, and Indian languagesbeing rich in inflectionaland
To build a training dataset, it typically requires a large man- derivative morphology, the roots of many of these words were
ual effort to annotate an example with the appropriate NSW- present in the training data. It suggests that we should use roots
category. For example, given a word corpus > 3M words of the context as the features to predict NSW-category, how-
in Telugu, Tamil and Hindi, we extracted 150-500K sentences ever, such approach needs morphological analyzers. Many of
containing an NSW. Annotating such huge set of examples the Indian languages fall into category of minority languages
needs lots of time and effort. To minimize such effort, we used where linguistic resources are scarce. Thus we wanted to inves-
a novel frequency based approach to create a representative ex- tigate sub-word units such as syllables and their combinations
ample set. as features for prediction of NSW-category.
NSW techniques uses context information for disambigua- Our experiments on POS-tagging on Hindi, Bengali and
tion with various window sizes, context information contains Telugu using syllable-level units further provided evidence that
a set of word like units which occurs in left and right side of syllable level features could be used as alternative and a first-
a NSW, and this information is to be considered as a features order approximation of root of a word [11]. After initial set of
characterizing a NSW. However, not of all context would be experiments to explore different possibilities of using syllable-
useful, so we used a window size of 2 (left and right) as a default level features, we confined to a set of following three syllable
and given to the pattern generator module. The pattern gener- level features. They are: 1) F1: previous ten and next ten syl-
ator takes the four tokens (two to left and two to the right of lables of a NSW, 2) F2: previous ten and next ten syllables and
a NSW) and generates 15 patterns using all possible combina- onset of each syllables and 3) F3: Onset, vowel and coda of
tions of 4(like examples, 0001, 0010, 0011, 0100, ., 111) where previous ten and next ten syllables.
1 represent presence of a token and 0 represent deletion of the Using decision forest, the final predictionof NSW-category
token. Given such 15 patterns for each example, these patterns is chosen based on voting on the outputs of the three decision
were sorted in the descendingorder of their frequency and based trees built using F1, F2 and F3. This strategy gets the results of
on a threshold a set of patterns were choosen and given to a na- each decision tree and performs a majority voting to predict the
tive speaker to annotate the NSW category. The user interface NSW-category. The performance of the decision forest based
was built such that if the native speaker couldn’t annotate the system using syllable level features is shown in Table 10. We
NSW with the given pattern, then an extended context was pre- found that the results of using syllable-level features for text
sented to him at varying levels. Using the frequency based ap- normalization performed significantly better than that of using
proach, we could reduce the training examples to around 1000- word-level features. This significantimprovement in the perfor-
1500 which a native could annotate within a couple of hours. mance is primarily due to syllables acting a first-order approx-
Having got the annotation done, we looked at level of context imation of roots of the context words and thus minimizing the
the native speaker has used to annotate a NSW. We found less problem of unseen context. The final performance of the text
than 10% of time the user has looked into a context information normalization system is further improved after using expander
more than a window size of two. module from 91.00%, 82.80% and 87.20% to 96.60%, 96.65%
and 93.38% for languagesTelugu, Hindi and Tamil respectively.
6.2. Performance of Base-line System
Using word level units and decision tree, we built a base-line
7. Conclusions
system to predict the category of a NSW. We have tested the This paper explained the nature and difficulties associated with
performanceof the system on a separatemanually prepareddata building text processing components of TTS systems in In-
obtained from a different source (web) referred to as Test-Set-1 dian languages. We have discussed the relevancy of font-

6th ISCA Workshop on Speech Synthesis, Bonn, Germany, August 22-24, 2007 192
identification and font-to-Akshara conversion and proposed a
TF-IDF based approach for font-identification. A novel ap-
proach of conversion from font-to-Akshara using the shapes
of the glyphs and the assimilation rules was explained. We
have also studied the performance of pronunciation models for
different features including contextual, syllabic and acoustic-
phonetic features. Finally we have shown that syllable-level
features could be used to build a text normalization system
whose performance is significantly better than the word-level
features.

8. References
[1] Hunt A.J. and Black A.W., “Unit selection in a concatena-
tive speech synthesis system for a large speech database,”
in Proceedings of IEEE Int. Conf. Acoust., Speech, and
Signal Processing, 1996, pp. 373–376.
[2] Black A.W., Zen H., and Tokuda K., “Statisticalparamet-
ric speech synthesis,” in Proceedings of IEEE Int. Conf.
Acoust., Speech, and Signal Processing, Honolulu, USA,
2007.
[3] Zen H., Nose T., Yamagishi J., Sako S., Masuko T., Black
A.W., and Tokuda K., “The hmm-based speech synthe-
sis system version 2.0,” in Proc. of ISCA SSW6, Bonn,
Germany, 2007.
[4] Sproat R., Black A.W., Chen S., Kumar S., Ostendorf M.,
and Richards C., “Normalizationof non-standardwords,”
Computer Speech and Language, pp. 287–333, 2001.
[5] HaileMariam S. and Prahallad K., “Extraction of lin-
guistic information with the aid of acoustic data to build
speech systems,” in Proceedings of IEEE Int. Conf.
Acoust., Speech, and Signal Processing, Honolulu, USA,
2007.
[6] Prahallad L., Prahallad K., and GanapathirajuM., “A sim-
ple approach for building transliterationeditors for indian
languages,” Journal of Zhejiang University Science, vol.
6A, no. 11, pp. 1354–1361, 2005.
[7] Garg H., Overcoming the Font and Script Barriers Among
Indian Languages, MS dissertation,InternationalInstitute
of Information Technology, Hyderabad, India, 2004.
[8] GanapathirajuM., BalakrishnanM., BalakrishnanN., and
Reddy R., “Om: One tool for many (Indian) languages,”
Journal of Zhejiang University Science, vol. 6A, no. 11,
pp. 1348–1353, 2005.
[9] Khudanpur S. and Schafer C., “http://www.cs.jhu.edu
/ cschafer/ jhu devanagari cvt ver2.tar.gz,” 2003.
[10] Choudhury M., “Rule-based grapheme to phoneme map-
ping for hindi speech synthesis,” in 90th Indian Science
Congress of the InternationalSpeech CommunicationAs-
sociation (ISCA), Bangalore, India, 2003.
[11] S. Chandra Pammi and Prahallad K., “POS tagging and
chunking using decision forests,” in Proceedings of Work-
shop on Shallow Parsing in South Asian Languages, IJ-
CAI, Hyderabad, India, 2007.

6th ISCA Workshop on Speech Synthesis, Bonn, Germany, August 22-24, 2007 193