Text to speech synthesis for bangla language

Published Online March 2019 in MECS

Software systems have become an inevitable part of our daily life. Nowadays, the usage of software is tremendously increasing day by day. With the demand for different kinds of software systems, text-to-speech (TTS) synthesis system has come forward. There are hundreds of areas where TTS systems are very much important such as robotics, warning system, alarm system, email reading, human-computer interaction and especially for people with visual impairment and dyslexia. Considering the necessity of such systems, many popular technological organizations such as Mattel, SAM, Atari, Apple, Microsoft Windows, Amaga OS, Texas Instruments TI-99/4A offer speech synthesis as a built-in capability [23,24].

A TTS system converts natural language text into speech and then, a computer system able to read text aloud. A speech synthesizer converts written text to a phonemic representation and then converts the phonemic representation to waveforms which can be output as sound.

There are several ways to create synthesized speech. Among them, concatenative synthesis and formant synthesis are very popular. Concatenative synthesis is based on concatenating pre-recorded speech of phonemes, diphones, words or phrases. Concatenative synthesis produces the most natural sounding synthesized speech because of its use of pre-recorded data. Formant synthesis makes the timbre of a voice or instrument consistent over a wide range of frequencies and generates artificial, robotic sounding speech.

In this paper, we are using a concatenative synthesis technique to generate natural sounding speech. Bangla is one of the most important Indo-Iranian languages which is the seventh most popular language in the world and spoken by a population that now exceeds 250 million [16]. Bangla is the primary spoken language in Bangladesh and the second most spoken language in India [4]. Several researches were conducted in Bangla speech synthesis but these are not enough to build a complete TTS system. Sometimes a large lexicon is necessary to design a TTS system which needs long processing time [6]. Bangla language has 50 alphabets and the English language has 26 alphabets, which is almost half of Bangla alphabets. Taking this into concern, we translated Bangla text to English to reduce the processing time and to be able to use the existing English phone set to generate Bangla speech.

There is some text to speech synthesis engines available nowadays. Among them, festival is an opensource extremely flexible concatenative TTS engine which uses diphones or other units to generate synthesized speech [21,25]. It uses Bangla lexicon to produce Bangla speech [1,6]. Festival is a large system with slow compilation process and high runtime memory requirement [2,22]. Flite is another speech synthesizer which considers the size and performance on embedded platforms that reduces its flexibility [7]. Considering the performance issue and flexibility, another speech synthesizer FreeTTS is developed based on the two speech synthesizers. FreeTTS uses algorithms of Flite and the architecture of Festival. It is found that FreeTTS runs two to three times faster than Flite [7].

This paper presents a Bangla text-to-speech synthesis system which is flexible, needs small processing time and produces a good understandable speech. Besides, we developed an intermediate XML based data representation feature which will help users to create speech based on their own dialect. It reduces to know the technical details to synthesize speech. To the best of our knowledge, ours is the first work on synthesizing Bangla speech using English diphone set that reduces the processing time for synthesization.

The rest of the paper is organized as follows: section II presents several existing works regarding text to speech synthesis system. Section III presents the proposed approach of text-to-speech synthesis system. Section IV discusses the experimental results. Finally, conclusions of this work and suggestions for future work are summarized in section V.

• II.    B ackground S tudy

Developing a text-to-speech synthesis system is a challenging task. There are many stages such as text normalization, text-to-phonemes conversion, prosodic emotional content detection, and speech synthesis are needed to accomplish to develop a complete TTS system.

Plenty of research works have already been proposed in Speech synthesis for different languages. Some early researchers tried to build machines to emulate human speech, long before the invention of electronic signal processing. In 1779 speech synthesis has come under the light when models of the human vocal tract were built that could produce the five vowel sounds (in International Phonetic Alphabet notation: [a], [e], [i], [o] and [u]) [8].

A pitch synchronous waveform processing technique for text-to-speech synthesis using diphones was presented in [19]. In this paper, several algorithms were reviewed in a common framework to improve the voice quality of a text-to-speech synthesis system. The framework was developed based on acoustical units concatenation technique [19,20]. A German text-to-speech synthesis system, MARY was proposed by Schröder, Marc, and Trouvain [17]. The systems main features are a modular design and an XML-based internal data representation. It allowed the user to access and modify the intermediate processing steps without having a technical understanding of the system. Though research in text to speech synthesization for western languages has reached in a good position but for Bangla language that is very few such as [1,10,11,14].

The work reported in F. Alam et al. developed a speech synthesizer for Bangla language [1,6]. This system is developed using diphone concatenation approach. It needs a lexicon with its pronunciation to produce speech. The lexicon contains ninety-three thousand entries [6]. The proposed system creates voice data for festival and additionally extends festival using its embedded scheme scripting interface to incorporate Bangla language support. It translates Bangla unicode text to ASCII according to Bangla phone set. However, there is no description of how the transliteration process works. Moreover, there is no description about letter-to-sound (LTS) rules developed for words that are absent in the lexicon.

Concatenative speech synthesis system based on Epoch Synchronous Non OverLap Add (ESNOLA) technique for Bangla text to speech synthesis is discussed in [10,11]. The ESNOLA algorithm is developed for concatenation, regeneration as well as for pitch and duration modification. Preprocessing module creates partnames database from the pre-recorded natural speech signals, text analysis module accepts input text and generates phoneme string and stress marker and synthesizer module generates speech through combining the slices of prerecorded speech.

PDF text to speech conversion process is discussed in [9] where other tried to analysis sentiment from Bangla text [3]. PDF represents different types of data as objects such as text object, image object and multimedia object [9]. The pdf to unicode text conversion process extracts texts from pdf objects and unicode text to speech conversion process produces speech.

Every language has standard and non-standard words. To generate speech all the non-standard words should be converted to their correct pronounceable form. There are several ways to identify and normalize non-standard words. Some researchers have identified several semiotic classes like text normalization [12,13]. Regular expressions were written in .jflex format to recognize each semiotic class. And a set of rules were used for tokenization and verbalization. Another approach used decision tree and decision list for disambiguation [14]. Though some works have been done in this domain, but still there are some problems which need to be accomplished to get a good quality sound.

• III.    M ethodology

To synthesize speech from text, we proposed a text-to-speech synthesis system for Bangla language. The overall architecture of the proposed system is given in Figure 1 where we have normalized, tokenized, romanized and synthesized the input text.

Fig.1. Architecture of Bangla text to speech synthesis system

• IV.    T ext N ormalization

A text document contains not only full words but also various other language units such as numbers, dates, symbols, and currency. While speech can be synthesized from full words directly (subsection III), all the other language units must be first consistently expanded into full words before they get synthesized. The language unit conversion process which takes place internally is called text normalization. Table 1 contains the list of language units along with their expanded format. In the following subsections, the normalization process of some language units is discussed.

• A.    Number normalization

Number is a mathematical notation to count, measure or label. Bangla numerals system has ten digits: ০ , ১ , ২ , ৩ , ৪ , ৫ , ৬ , ৭ , ৮ , ৯ like Hindu–Arabic numeral system [15]. There are hundred numerals started from zero (0) to ninety-nine (99) (table 2). For numbers above 99 there are five main systems for naming numbers in Bangla (table 3).

If a text has only digits ( ০ - ৯ ) or digits separated by a comma (“,”), then it will be recognized as a number. After identifying the language unit as a digit, the following procedure converts a number to its pronounceable form.

• 1.    Firstly, Bangla number is converted to the English number. This process works by replacing Bangla digits to the corresponding English digits. And, the relationship between Bangla and English digits is: ০ ->0, ১ ->1, ২ ->2, ৩ ->3, ৪ ->4, ৫ ->5, ৬ ->6, ৭ ->7, ৮ ->8, ৯ ->9 .

• 2.    After converting a number from Bangla to English, it is checked if the number is less than “ ১০০ ” (100). If the number is less than all number units, then the number’s corresponding pronounceable form is taken from Bangla numerals (table 2). But, if the number is greater or equal to a number unit (descending order), the number is divided by that unit and the quotient and remainder is calculated.

• 3.    The calculated quotient and remainder are checked whether it is zero or not. If quotient or remainder is not equal to zero, it is passed again to process 2. And the units Bangla pronounceable form is added to the pronounceable text.

For example, “১০০৩৩২” is a Bangla number and converted to English number 100332. 100332 is greater than number unit 100000. Therefore, 100332 is divided by 100000 and the remainder is 332 and quotient is 1. The quotient is not equal to zero and less than 100, therefore, the pronounceable form now is “এক লক্ষ” (one hundred thousand) (quotient + units pronounceable form). However, the remainder is greater than unit 100. Therefore, the remainder is divided by 100 and again the quotient 3 and remainder 32 is calculated. Now, both are less than 100, therefore the pronounceable form of the quotient is “িতন শত” (3 hundred) and the remainder is “বত্রিশ”  (thirty-two). Finally, “১০০৩৩২” will be pronounced to “এক লক্ষ িতনশত বত্রিশ ”.

• B.    Date normalization

  According to the national and official Calendar of Bangladesh, the date format is “ দদ - মম - বববব ” (dd-mm-yyyy). A text is identified as a date unit, if it contains a one to thirty-one, following a separator and a one- or twodigit number denoting a month ranged from one to twelve with the same separator and a two- or four-digit number denoting a year. People also use some other types of date formats like “Day number – month name – year” for example “ ২ জলু াই ২০১৭ ” (2 July 2017). These types of dates can be one- or two-digit number denoting a day ranged from identified if the text contains a one- or twodigit number, a separator, a text denoting the month and a four-digit number denoting the year sequentially.

Table 1. Language units with their expanded format

Language unit	Non-standard format	Expanded format
Cardinal number	১১৩০	a^lwy a^*s Sy*
Fractional number	১০৫ . ০২	a^*s яр y^f^ жчч жчч h#
Ordinal number	১ম , ২য়	yyyy , yylSly
Date	০২ / ০৫ / ২০১৬ or ২ মম ২০১৬	H# 4L H# W-й ЧИИ
Phone number	+ ৮৮০১৮১৩৩৭৫১৮২ or + ৮৮০ - ৯৬৩৯৫৩৭১৯০	yyisy^ ж6 ж6 жчч a^ жё a^ 1^^ fiy yfs Яр a^ жё h# or or ^^is^^ Жё Жё ж^ч чу ^y fw^ yy ^fp чТь f^4 yis a^чу жчч
Range	১০ - ১২	уж 4L^t УШ
Roman numerals	I, II, III, IV, V	я^^^ , yylsly , sSly , ps^;^ , ^^p4 ..
Time	১২ : ৩০ : ১৫	yiy6i syl*, У1ч1ё ясчгуг ЯМШ1 чс^сч»
Unit and measurement	°, ’, ”, %	®fyy1 ? yfyf6 , чс^сч» , УТ7?ж

There are four date component separators which are the followings.

• 1.  “/” – stroke (slash)

• 2.   “.” – dots or full stops (periods)

• 3.   “-” – hyphens or dashes

• 4.  “ ” – spaces

After identifying the date, we have converted it to its pronounceable format. We have separated the date unit as day, month and year.

The day is expanded using the number normalization algorithm. Bangla calendar has twelve months in a year like English calendar system. If the month is a text like “^тИ” (July), it remains as it is. If the month is a number, we replaced the number by the corresponding month text. For Bangla year there are two formats to pronounce a year. If the year is less than one thousand or not thousands such as 1YXX, 2YXX ... (here, x represents any digit and Y represents a digit not equal zero), it is pronounced by grouping.

Table 2. Bangla numerals

o	ЧШ	bb	4ЖЯ1
b	4^	bS	314
S	4t	b®	C®s
®	f^	b8	Cbly
8		b6	^C43
6	'Ш	b®	C314
®	S^	b8	У®«
8		bb	^^ИЯ1
b	^6	bb	tim
b	43	So	f34
bo	4^	Sb	4^4
...	...	...	...
b8	gylyyy#	b8	yl®lyyy#
b6	УЫУУУ#	bb	^ыууу#
b®	^lyl^yy#	bb	yfdHyy#

Table 3. Units for naming Bangla numbers

Number notation	Power notation	English numbering unit	Bangla numbering unit
Ьоо	102	One hundred	®^ Ч®
Ьооо	103	One thousand	5^ ГТ^Ш
Ьооооо	105	Hundred thousand	5^ 3 ^
Ьооооооо	107	Ten million	а^ C^116

Table 4. Date normalization

Regular text	Identified date	Normalized date
b ^3lS sob® ®1Ж^ ^y^^ ^y^^ Sly^l^lyl ЕТ^ТУ ^S ^f^ ^16^1^ c<№Ri® ^f^^^cT ^СЯ 1	b    g^iS sob®	5^ ^ffllS и# гТ^ТУ ьТ^ТУ У! 83
S / ® / Sob8 sryf^t ^Т?ЗТ^ГЧ ^yl^tB из ^yi^^^l y^y ^ytl	S / ® / Sob8	и# УШ и# гТ^ТУ 31У

The last two digits make a group and the rest digits make another group and the word “ C^1 ” (sho) is added to the last group. For example, 1971 will be pronounced as “ @f^^C^I 4№^sT (unissho ekattor). And, if the year is like 10XX or 20XX, it is converted using the number conversion algorithm. For instance, 2012 will be pronounced as “ ^t S1^1S ^Ш (two thousand twelve).

• C.    Currency normalization

The Bangladeshi taka (Bangla: ' 61^1 ' ) is the currency of People’s Republic of Bangladesh and its sign is ` ৳ ' . There are two ways to represents currency or amount of money. Firstly, the currency sign ` ৳ ' which comes before a number like ' Ъ Ьоо ' . Secondly, when the word “ 61^1 ” (taka) comes after a number like “ Ьоо 6ш ” (100 taka). The currency unit needs to be normalized to the corresponding pronounceable form for both of these situations. The correct pronounceable form of ' Ъ Ьоо ' or “ boo 61^1 ” is ‘W 4® 61^1 ”.

To recognize currency from the text, we have created two currency recognition formats like:

• •    “Ъ - Space - N” or “Ъ - N”

• •    “N - Space - “ 61^1 ”

Here, N refers to a number. The number may have a comma (“,”) to separate special units (“^ , ^1^1Я , ^^ , №116” ).

We have used the same algorithm to normalize currency which is used to normalize number. After recognizing a text as a currency unit, we separated the word “ 61^1 ” or the currency sign ' Ъ ' and get the number. Then, the number normalization algorithm generates pra onounceable form of that number.

Finally, the word  “61^1”  is added after the pronounceable text.

D. Phone number normalization

A telephone number is a sequence of digits assigned to a fixed-line telephone subscriber station connected to a telephone line or to a wireless electronic telephony device such as a radio telephone or a mobile telephone or to other devices for data transmission via the public switched telephone network (PSTN) or other private networks. The subscriber phone number in Bangladesh is a unique 11-digit long number. The country calling code for Bangladesh is +880.

The typical format for a mobile phone number is: “+880-1X-NNNN-NNNN” and typical format for a telephone number is: “+880-96XX-NNNNNN”.

For mobile and telephone number, +880 is the a country code, X is operator code and N is subscriber number.

When dialing a Bangladesh number from inside Bangladesh, the format is:

• •    0 – operator code (X) – subscriber number (N) or

• •    96 – operator code (X) – subscriber number (N)

If a text has +880 following 1 or 96 and an eight-digit number, it will be recognized as a phone number. When dialing inside Bangladesh, the country code is not necessary. In that situation, if a text has 1 or 96 following an eight-digit number, it will be identified as a phone number.

Phone number is actually a sequence of digits. After identifying a text as phone number, the digits of that number are replaced with their corresponding pronounceable form.

• V.    T okenization

Tokenization is the process of demarcating and possibly classifying sections of a string of input characters. In tokenization, a given character sequence is chopped into pieces called tokens. A token is an instance of a sequence of characters in some particular document that is grouped together as a useful semantic unit for processing.

Table 5. Bangla alphabets romanization

Grapheme category	Bangla grapheme		Romanized form
Vowel	Vowel	Vowel mark	-
	^	-	o
	^	oT	a
	5	fS	i
	^	St	i
	^		u
Consonant ( ^Ч^йт^Ч^Ч )	^		K
	Ч		Kh
	я		g
	4		gh
	^		ng
Consonant conjuncts (WWI)			k…k
	46		nt
	ИЧ		dh
			kkho
			cch

Tokens are identified based on the specific rules of the lexer. Some methods used to identify tokens include: regular expressions, specific sequences of characters termed a flag, specific separating characters called delimiters and explicit definition by a dictionary. Special characters, including punctuation characters, are commonly used by lexers to identify tokens because of their natural use in writing. Like English, Hindi and other South Asian language, Bangla language also uses whitespaces to tokenize a sequence of characters into individual tokens.

In this paper, the punctuation characters are used to tokenize sentences and then the sentences are further tokenized to words by a whitespace character.

• VI.    R omanization

Romanization is the representation of a script in Latin script. Bangla is a segmental writing system and its graphemes represent the phonemes. Bangla Script has 11 vowel graphemes and 39 consonant graphemes and more than two hundred consonant conjunctions. We have Romanized Bangla script according to Bangla grapheme set. Table 5 shows Bangla graphemes sets for vowel, consonant and consonant conjuncts with their corresponding Romanized from and table 6 shows the romanization process.

Table 6. Bangla word romanization process

Bangla word	Bangla syllable	Corresponding English syllable
жчтч	ж + ч + T + ^	a + m + a + r
И1Ч	И + I + ч	d + e + s
ЧTS^TИIЧ	^ + T + ^ + ^ + T + И T + И + I + ч	b + a + ng + l + a + d + e + sh
^^f^	^ + ^ + Г+ ч	ro + h + i + m
^H^fri	^ + и + ^ + + ^ ^	k + o + k + i + l

Based on the romanization process, each token is romanized to Latin scripts. We have designed romanization rules based on vowel and consonant combinations. Some of the rules are described below:

• 1.    The vowels are romanized directly according to its corresponding romanized form (table 5).

• 2.    If a consonant is in the last position of a word, it is replaced according to its romanized form (table 5). For example, in the word “ ^^^^ ” (bokul) , “ ^” is a consonant which is in the last position, so ’l’ replaces the letter “ ^”.

• 3.    If a consonant is not in the last position of a word and if there is no vowel after it, ‘o’ is added along with the consonant. For example, “^fl ”   is

• 4.    If the character “ o ” is found in a word, the letter before and after it is taken into consideration to make a consonant conjunct and then the conjunct is looked in the Bangla alphabets romanization table (table 5). If the consonant conjunct found, its

• 5.    If the character “ os ” is found in the middle of a word, the consonant after “ oS ” is placed in its position. For example, in the word “^f is pronounced as “ ^y^ ” and its romanization form is “Dukkho”.

romanized as “rohim”. Here, ‘r’ is for “ 3 ” and ‘ o ’ is added after it to make the pronunciation correct.

corresponding romanized form replaces it. If it is not found “ о ” is escaped.

• VII.    S peech S ynthesis

Synthesized speech is the ultimate production of a TTS system. The text is converted to phonemes based on the phonemes database. Phoneme is the fundamental unit of sound in a language. Then the prosody analysis analyzes the prosody of the phonemes, words, and sentences to determine the appropriate prosody. The prosody and phonemes information are used to produce audio waveforms of the sentences.

In this paper, to synthesize speech, we have used MBROLA voice [27]. It is a 16 kilohertz (kHz) male voice. After romanization, the text is converted to an interface named FreeTTSSpeakable. The source text that need to be spoken is first converted to it. Then the FreeTTSSpeakable is sent to voice interface which is the central processing point of speech synthesis. It takes FreeTTSSpeakable as input, converts the it into a series of utterances using the MBROLA voice and generates audio output.

• VIII.    X ml D ata R epresentation

Along with speech synthesis system, we have developed an xml-based data representation system depending on the language units. Each language unit is given a specific tag. After tokenization (subsection II), all the tokens are given their corresponding language units tags such as date, time, word and currency. We have used regular expression for the purpose of langauge unit identification. This is an intermediate data representation where users can modify the data representation without knowing the technical details of the system and can generate speech based on their own dialect from the xml.

Besides, users can parse data from the xml data tree (Fig.2). For example, if a user wants to get all the date, by parsing the xml s/he can get all the date.

Fig.2. XML data presentation

• IX.    E xperimental E valuation

To produce output speech, input text is taken from different sources. The sources of input texts are daily newspaper, poem and short stories. We considered the most popular daily newspaper prothom-alo, the famous poem of Bangla literature “ ^T4T ^^fa ^T ” and Bangla short story “Chuti” (“ gfS”) [27], written by the Nobel laureate Rabindranath Tagore. In table 7, the input text along with its corresponding romanized form is shown.

To get the result, we have selected two groups of people. One is graduate students and the other group of people is senior citizens. We let them to hear the produced speech and write down the words they have heard. Moreover, result shows that graduate students are more attentive and understand more words than senior citizens. The result shows that graduate students understand clearly 68% of the produced speech where senior citizens understand 60% of the produced speech (Table 8). Senior citizens understand less of the produced speech, because of their physiological aging and changes in cognitive ability [26]. Moreover, result also varies based on the sources where the best result is found for poem and the least result is found for short stories. The average accuracy for newspaper, poem and short stories are 67.37, 71.87 and 64.52 for graduate students, and 59.6, 62.5 and 59.6 for senior citizens.

Table 7. Input Bangla text with corresponding romanized text

Bangla text	Corresponding Romanized text
ути ® утусу 1 уиТууГ аг иВГ ^утит умим у? иу , ьиус ФЫ«с УТУСИ У.Ы-с£ 1 УТ^1И УУ^С У1И1 ® УТУМ аг ибТ аУУ УС УИСУ У? у? иу । ути у? «tyf уус ууууэдГ । ууусу ^ум учи ууТ^ису ииу иТус УТУ , «УИ У1У - УУ^ - И|У — а^ «ТИ У? ИСУТ УТУ 1 «Т УИСУ УИсУ ^Усг УУИТЧ уус^си <Т^^|У1 уу|у ^г^|У ИТФбИ 1 "УУТУ ^УТУ УТ«У®1 УМУТУТУ УУ« УМИМ У|У^ УТ УМУ» ^МУС ИУ«С УТУУ ИСУТ УТУ УУ у?# ФиТ® - УТИТ ауб У% ЬЫУС УУУС 1 УШ^ТИЧУТ УУСИ ^УМУ уиууч1«Г^ум УТУМ 1 ^УТУ РтуУису УТ£с УУ у? иТУТУс «ЕуГ ЭД УТУМ , УТИТ «М УУТИ«У 1 УУ ЭД аг аг УТИТ ® УТУМ аУУ У? yy 1 иВМ УИСУ УМУ УУУУУ 1 «тг аг аг ибГУМУ УУ«ТУ ^У«М УИСУ усиТ । Уису Уису Фуууу , Уису УИСУ "МУСУ УТ УИ««1У УУУТИ УУТ у|у УТИТ - УТУМ иТУ 1 уиЬис УИЬТИС ИТИТ «МУФТИ , иТИ уТ?У1 МУМ^ИС «УСУУМ® , УТУТУУМ» УТУТУУМ» УТУУС® МИТИСУ исис и1уиТУВ«тус ус ууу у1% ИСг 1 «у Фуууу , иТису ^У^ УУС ^ИУТ^ у1и« умитусу у? ЫУ ууГ 1 УУТ 1 ^т«1у имиу иТуу уТ?ут яуис УСУУУТУТУ УМИМ ^УМ^ИС ^УМ^ИС УСУС УТИТ - УТУМУ УТ^УС УУ ЯУбТ ^ИУТ утг ИТ 1 аУУСУ ауису УУИИ УУЭДС УТ ^МУСУ уу^т«усуТ«с УТИТ - УТУМ у^иту , УТУ УТ — УТФУС УУС иТ«С ^У ИТ1 УС У|Щ1«С УМЖУ ибит ибс^с , ибс^с , УСУТИТС® ИСУТ УТУ УУ^И РТУТИМ ИТИУСУТ иижу уутэдТу У1УУ1 УТИУС иТ^СУ УУИСУ УМУТУбГУИУС УТИТ УТ УТУСИ УТУТУ УУС уус итИУ ^ь^си исуу|«уТУ 1 ^уус ууТУси - уУУиТ^Т' , ^усу - УИ^«ГУУС ИГь^С ®г УИУС У^^М У? ИТУУС УУИС 1	sada o kalo. jodi boli ai duti alada kono rong noy , chomoke uthote paren onekei. biggan boloche sada o kalo ai duti akok kono rong noy.,sada rong toiri kore surjroshi. Surjer alo jokhon prijomer moddho diye jay , tokhon lal sobuj nil ai tin rong dekha jay.,ta onek agei prman korechen biggani sjar aijak niuton. kono kagoj ba bord jore ghurote thakole dekha jay sob rongoi udhao , sada akota kichu chokhe poroche. bigganira bolen alor onuposthiti holo kalo. abar iosolpider kache sob rong misiye toiri hoy kalo , sada to kanovas. bola hoy ai sada o kalo akok rong noy . dutoi onek ronger somonnoy. tai ai duti ronger prvab hoyoto onek besi. bises bises upolokkho , bises abeg ba onuvutir prkas kora jay sada kalo diye. poschime nana onusthan , din kingoba ayojone dresokod , kalarokod thakoleo amader dese nirdistvabe se rokom kichu nei .,tobe upolokkho , diner aboh bujhe amorao kintu posaker rong thik kori.,jatiy sok dibos kingoba akuse februyarir kono ayojone gele sada kalor baire khub akota amora jaina .,akuser prthom prhore ba vorer prvatoferite sada kalo posak , khali pa kauke bole dite hoy na.,je barite soker ghotona ghoteche , sekhaneo dekha jay sojon harano manusera soker prathomik dhakka samole nijer poroner posakoti bodole sada ba kalo peaosak pore samil hocchen sesojatray., asole poribes poristhiti , abeg onuvuti bole dicche oi somoye kon rong thakobe porone .
^ ИСУТ УТУ «ТУ У^ , ^ ГИТИСУ УТ , ®г УТИС«С УТУ УУС УТИТ умуТУ усиуТУ ^т । ® усиуТ «г уту уТ , УТИ«т «т« ь|у уГ , аубт уиГуту УИИТ УУС УТУУ УУУ УТУ 1	oi dekha jay tal gach , oi amader ga , oi khanete bas kore kana bogir cha . o bogi tui khas ki , panta vat chas ki , akota jodi pas omoni dhore gapus gupus khas .
утууиТУ|су ууиту ^ВТУ ьуууу«1у итуту ьб ууТУт аубт и«и «ТУМИУ $$У 1 ИИ1У УТУС аубт УУУТУ« ИТУУТуЬ ИТУ«УС уу|И«уТ« ^§уту уу«1уиту у«ТУт ^ТУ i yyfy ^#у , усВ ууу; ИТУТУТ y«t#yt У^У УТ#УС| уиТУ ^^У , стбт УУУ ИТУТУ Т У®Т#УТ У^ут УТ^УС| УС УУУ«ГУТЬ , ^УУУУ УТУС «Т^ТУ УС у«утиГуТУ^иу УУуу у1Уу«Г ау? ууУУт усни ^ус , «т^т# ФууууТ уУУ УТУУСУТ а УУУ«ТУС УИУУИ УИИМИИ ууТУ 1 умиу утиТУт ууус# ууи иис^тусуусу y^f« утус ууу® ^#уту Фуууу уу1«с^с аии уиус уВТусу уиТУЫ итуиуту уи«1у®тус ус# у«1У Фуус уТУ 1 ууТУ ууТУ 1 ^сусут «т^ту а^уу Фиту ^иту1иу исуТу т уТ^ уТиуу ^#у ^У УСУ 1	balokodiger sordar fotik chokrbortir mathay chot koriya akota nuton vabodoy hoilo . nodir dhare akota prkando salokastho mastule rupantrit hoibar prtikkhoay poriya chil . sthir hoilo, seta sokole miliya goraiya loiya jaibe. je bektir kath , abosok kale tahar je kotokhani bishoy birokti abong osubidha bodh hoibe , tahai upolobdhi koriya balokera a prstabe sompurn onumodon koril . komor badhiya sokolei jokhon monojoger sohit karje prbritt hoibar upokrm koriteche amon somoye fotiker konistho makhonolal gomvirovabe sei gurir upore giya bosil . chelera tahar airup udar oudasinj dekhiya kichu bimors hoiya gelo.

Table 8. Result evaluation

Testers category	Testers	Correctly written words from total words			Percentage (%)
		Newspaper - 220 words	Poem - 32 words	Literature - 85 words	Newspaper	Poem	Literature
Graduate student	Student 1	147	23	53	66.8	71.8	62.3
	Student 2	155	25	57	70.5	78.1	67.1
	Student 3	141	21	51	64.1	65.6	60
	Student 4	137	20	55	62.3	62.6	65.9
	Student 5	154	24	55	70	75	64.7
	Student 6	155	25	57	70.5	78.1	67.1
Senior citizen	Citizen 1	127	22	48	57.7	68.8	56.5
	Citizen 2	133	21	55	60.4	65.6	64.7
	Citizen 3	122	19	45	55.5	59.3	52.9
	Citizen 4	129	20	51	58.6	62.5	60.0
	Citizen 5	127	23	50	57.7	71.9	58.8
	Citizen 6	130	15	55	67.7	46.9	64.7

• X.    D iscussion

The produced speech of the system is understandable but has lacking in its naturalness. In this system, we have used MBROLA diphone database, because developing a diphone database for Bangla language itself is a lengthy process. We have produced Bangla speech using existing English diphone database which reduces lots of task to synthesize speech. English language has 26 alphabets and Bangla language has 49 alphabets.

The English diphone database is small considering Bangla diphone set and it reduces the memory requirement of the system.Therefore, we can say that the computation time will be short considering the small diphone set in English language. The result of our proposed system is satisfactory as we have achieved this result with a very lightweight system without creating any new Bangla lexicon or a diphone database.

• XI.    C onclusion and F uture W ork

TTS system has become popular due to its use in various sectors. But Speech synthesis technique for Bangla language is not very satisfactory. This paper introduces a new lightweight TTS system for Bangla language which uses existing English diphone sets to generate Bangla speech. The proposed system produces good quality understandable speech in small processing time. However, the sound quality is not very natural. To generate good quality natural speech, our future focus is to explore other state of the art techniques.

A cknowledgement

This research has been supported by The University Grant Commission, Bangladesh under the Dhaka University Teachers Research Grant No-Regi/Admn-3/58763, dated on 21.03.2017.