Concatenative Speech Synthesis for Amharic

Using Unit Selection Method

Eyob B. Kasie Yaregal Assabie
Transnational Computer Technology Department of Computer Science
Addis Ababa, Ethiopia Addis Ababa University
Addis Ababa, Ethiopia
[Link]@[Link] [Link]@[Link]

ABSTRACT minimizes the need of digital signal processing by acquiring

In this paper we propose algorithms and methods that address multiple phones for the same phoneme. Unit selection based
critical issues in developing a general Amharic text-to-speech synthesizers choose suitable fragments from a database of speech
synthesizer. Converting grapheme to phoneme in Amharic is a recorded from a speaker and join them together with minimal
very challenging task because of the two necessary and yet signal modifications. Minimal modification of the speech signal
orthographically unrepresented components of the language – produces highly intelligible and natural sounding utterances
epenthesis and gemination. Modeling prosodic features of various instead of buzzy or robotic speech. Amharic is suitable for unit
speaking styles is also the other challenging task in developing selection approach because of its phonetic nature [10]. The one to
Amharic TTS. Making use of orthographic property of verbs in one correspondence between orthography and phonemic
their perfect form, this work introduces set of rules that can be representation has much value in regard to solving one of the most
used to locate phones that need to be stressed and algorithm that critical problems in unit selection approaches – the discrepancy
inserts epenthetic vowel wherever necessary. This paper also between phones asked by the front-end and selected from the
presents methods to represent intonations that vary according to sound bank [3]. Sebsibe et al [9] has shown the results found in
punctuation marks (“?” and “።”) and phoneme location. In using the Festvox framework for building unit selection voices for
addition to these issues we also introduce transliteration rule for Amharic language are encouraging. Tadesse et al [10] have
numbers and phone labeling scheme that captures allophonic claimed highly intelligible speech by integrating a morphological
variations in regard to gemination, epenthesis, and declarative analyzer called HornMorpho to locate geminates and epenthesis.
versus judgment request intonations. The system is evaluated for As far as we know, no work has been done on representing
naturalness and intelligibility by ten fluent speakers of the Yes/No and Declarative prosodies. Amharic, being a Semitic
language and experimental results are reported. language that has many speakers next to Arabic [7], is worthy of a
thorough study in the area of natural language processing. In this
study we introduce rules and methods useful for developing
Categories and Subject Descriptors general purpose TTS for Amharic. Amharic language issues such
I.2.7 [Artificial Intelligence]: Language Generation, Speech as epenthesis, gemination, declarative versus interrogative
Recognition and Synthesis, Text Analysis prosody modeling are studied and algorithms are developed to
address them. Allophone labeling scheme is introduced for easily
locating target sound units which addresses one of the major
General Terms difficulties of the unit selection approach, which is selecting the
Algorithms, Design, Experimentation, Human Factors, Languages right unit from multiple sound units. Sound units are recorded and
segmented in such a way that each sound unit captures the
Keywords relevant allophonic variations.
Unit Selection, Concatenative Synthesis, Amharic TTS,
Gemination, Epenthesis, Sentence Prosody
2. LINGUISTIC FEATURES OF AMHARIC
Amharic is a Semitic language and the official language of the
1. INTRODUCTION government of the Federal Democratic Republic of Ethiopia
Speech synthesis is the ability of a computer to produce spoken (FDRE). The Amharic orthography, as represented in the Amharic
words. Text-to-Speech (TTS) synthesis systems take text as input Character set - called ፊደል (fiidel) 1 consists of 276 distinct
and generate acoustic signal as output. Synthesized speech can be symbols. In addition, there are twenty numerals and eight
produced by several different methods. This work uses punctuation marks [10]. There are thirty three (33*7=231) core
concatenative’s special approach known as unit selection that orthographic symbols, each of which has seven different shapes,
usually known as orders, to represent the seven vowels. Each
consonant and the seven vowels in combination represent
Permission to make digital or hard copies of all or part of this work for consonant-vowel syllables [12]. Each of these consonant and
personal or classroom use is granted without fee provided that copies are vowel graphemes can appear independently or can form a single
not made or distributed for profit or commercial advantage and that letter.
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee. 1
MEDES’12, October 28-31, 2012, Addis Ababa, Ethiopia. We have used the transliteration scheme introduced by [9] for
Copyright © 2012 ACM 978-1-4503-1755-9/10/10...$10.00. converting Amharic scripts to their Latin equivalent.

- 27 -
2.1 Epenthesis production. The following figure (Figure 1) shows the architecture
Epenthesis is the process of inserting the epenthetic vowel of the system.
between consonant clusters. Epenthetic vowel is a high central
vowel which Amharic language uses to break up consonant Input text
clusters. This vowel is phonetically represented as ‘ɨ’ or ‘ə’. In Text-Preprocessor
Amharic, consonants (represented as the sixth order characters in Delimiters Input Text
Tokenizer
the series) may or may not be followed by this vowel [4].
Numeral Transliterator
Mulugeta’s [6] work on epenthesis in Amharic language can be Transliterator
Transliteration
summarized as - consonant clusters are allowed only in medial List Abbreviation Transliterator
and final positions of a word as long as the cluster is of size two;
and consonant clusters of size two are allowed at word final Tokenized & transliterated words
position if the sonority of the penultimate is equal to or greater Grapheme-to-Phoneme
than that of the last. In case of clusters with three consonants,
Phonetic Labeler
epenthesis occurs between the second and the third consonant.
Epenthesis Controller Gemination Controller Sentence Modeling
2.2 Gemination
It is the other critical factor that determines pronunciation and Phone set
semantics in Amharic language. Like epenthesis gemination is Sound Generator
also not shown in the language’s orthography. However, unlike
Sound Sequencer
epenthesis, developing a set of rules for gemination in Amharic is
an intricate issue. Tadesse, et al. [10] gives two broad foundations
in order to determine the gemination of consonants: lexical and Sound-Units Sound Player
Sound
morphological. According to their work, lexical geminations
cannot be easily predicted without making contextual Figure 1. System architeture of the Amharic speech synthesis.
disambiguation. However, morphological geminations can be, in
most cases, predicted by observing the orthography of the word. 3.1 Transliteration of Numerals
This prediction is especially successful on verbs. In Amharic, Even though old/traditional Amharic texts use the Geez numerals
almost all root verbs geminate either the penultimate or the final for representing numbers, these days, however, the trend has
radicals in the perfect form. The obvious exception from this rule changed so much that virtually all Amharic writings follow the
is the bi-radicals (which are actually tri-radicals underneath) that Arabic numerals for representing numbers. We will follow that
delete their middle consonant or biradicals that do not begin with trend and propose a transliterating scheme for the Arabic
a vowel [6].The gemination pattern for tri-radical roots in their numerals. A relational table is used for the scheme. Table 1 shows
perfect form is to geminate their penultimate consonant and for bi- the design and sample of the database table. Each digit from 0 to 9
radicals, it is to geminate their final radical. Anne and Klaus [1] is represented by 15 rows according to the power position it
further affirm that affixes are of no influence that the penultimate assumes, which means the transliterating scheme works for up to
radical maintains its germination during derivation. However, this 999,999,999,999,999.
seems to work only if the affixes do not change the tense of the Table 1. Transliteration of numerals.
verb, i.e., the tense stays in its perfect form.
Digit Power Position Value
2.3 Prosodic Differences of Declarative and 5 0 amst
Interrogative Sentences 5 1 hamsa
Ralf Kompe [8], summarized intonation differences among 5 2 amst meto
statements as the following: a falling intonation (pitch) marks an 5 3 amst sxi
utterance as a declarative, a rising contour of pitch indicates a
question, and Continuation-rise signals that there is a (prosodic) 3.2 Grapheme-to-Phoneme Conversion
clause boundary, but that the speaker has not finished his turn yet. One of the most difficult jobs of unit selection synthesis is
Generally speaking, major differences of intonation are especially choosing the appropriate phone from multiple phones. And this is
vivid at sentence-end positions. We did not find literatures that largely determined by the grapheme to phoneme conversion
researched prosodic differences between declarative and technique one employs. Phonemes differ according to
interrogative sentences in Amharic. But as noted by Joyce context/position, part of speech, and neighboring punctuation. In
McDonough [2] prosodic differences of declarative and the sections that follow algorithms/techniques are shown to
interrogative sentences are universally applicable. He says - “the address epenthesis, gemination and declarative-interrogative
distinction between declaratives and interrogative and focus prosody.
constructions is among the most universal characteristics of
intonation. Later in this paper we show waveform analysis
3.2.1 Epenthesis
Epenthesis is orthographically unrepresented aspect of Amharic
performed on declarative and Yes/No question statements.
phonology that plays a critical role in synthesizing speech.
The following algorithm (see Listing 1) is proposed to insert the
3. PROPOSED SYSTEM epenthetic vowel ‘ɨ’ wherever appropriate with the exception of
In this section we will discuss the algorithms and/or the methods gemination. The sonority level of consonants is as follows:
proposed on three important phases of a speech synthesis system – Glides>Liquids>Nasals>Fricatives>Stops
transliteration, grapheme-to-phoneme conversion and phone set

- 28 -
 extensive use of lexica and contextual disambiguation; second,
1. Split the given word into character array
almost all types of verb roots, uniformly, geminate their
2. While there is character in the character array
penultimate radical in the perfect form; third, it is easy to
2.1. Check if the character is consonant (which means it is
determine the location of Amharic verbs in any simple Amharic
the 6th order in the alphabet series)
sentence. One other limitation that should be explained is why
2.2. If the position of the consonant is word-initial, insert
only the person, gender and number inflection forms are chosen.
epenthetic vowel
This is because these inflections are represented by adding only
2.3. If the position of the consonant is word-final, ignore
suffixes; and according to [1], suffixes are of no influence to the
epenthesis
2.4. If the position of the consonant is next to word initial general gemination rule. Prefixes such as “Aስ” (as) change/add-to
and is not the penultimate character, ignore epenthesis the gemination rule because of assimilation processes. For
2.5. If the consonant is neither the second nor the example the word “ዘረጋ” (zeregga) will geminate “ዝ”(z) when
penultimate character, check its adjacent neighbors and prefixed with the bound morpheme “Aስ”(as) because of the
insert epenthetic vowel next to it if it is the second similar consonants “ዝ”(z) and “ስ”(s). We have developed an
consonant in a consonant cluster of three; else ignore algorithm (see Listing 2) that lists orthographic-based rules to
epenthesis geminate consonants of verbs in their perfect form and are only
2.6. If the consonant is the penultimate character and the inflected according to person, gender and number.
final character is also a consonant then check their
sonority level; if the sonority level of the penultimate is 1. If the verb is bi-radical and ends with “ጨ” (cxe) then “ጨ”
less than that of the final then insert the epenthetic (cxe) is geminated.
vowel after it. 2. If the verb is mono-radical and starts with vowel then the
3. End while final radical is geminated.
Listing 1. Algorithm to control epenthesis. 3. If the verb is bi-radical and the penultimate radical is first
order and the final is fourth order then the final radical is
geminated.
3.2.2 Gemination 4. If the verb is bi-radical and starts with a vowel, step 3
In this section we propose an algorithm that determines applies. But if the final radical is first order instead of being
geminations from orthography on Amharic verbs in their perfect fourth order, the penultimate is geminated instead of the
form. The algorithm makes consideration of inflection of the final.
verbs in regard to person, gender and number. An exhaustive list 5. If the verb is tri-radical, and all its radicals are first order, the
of suffixes used for person, gender, and number inflections is penultimate is geminated.
shown in Table 2. 6. If the verb is tri-radical (excluding the suffix) and it has one
Table 2. Suffixes of Amharic verbs based on person, gender of the suffixes listed in Table 2, the penultimate is
and number inflections. geminated.
7. If the verb is tri-radical or bi-radical that starts with a vowel
Suffix Suffixes and has none of the suffixes listed in Table 2:
Length
7.1. The final radical is geminated if its penultimate radical
One n, nx, sx, t, w, k, h, hu, ku, c is first order and its final is fourth order
Character 7.2. Both the final and the penultimate radicals are
Two sxnx,cnx,ken,cn,kusx,chu,ch,nh,csx,nsx,kuat,new, geminated if both are fourth order and represent
Characters huat,cat,kut,sxat,nat,cew,enx,kenx,kunx,en,wuh,hu different phonemes1.
sx,hut,huh,let,bet,lat,bat,bh,lh,lsx,bsx,lenx,benx,ln, 7.3. The final radical is geminated if both the final and
chunx,kuachu,chun,nachu,cachu,kuacew,sxacew, penultimate radicals are fourth order and represent the
nacew,cacew,huachu,huacew,lacew,bacew, lachu, same phoneme
bachu,cbet,clet,cbat,clat,cbh,clh,clsx,cbsx,cbnx,cln 8. If the verb has more than three radicals and it ends with one
x,cbn,cln,kbet,hbet,klet,hlet,hbat,hlat,kbat,klat,kbn of the suffixes listed in Table 2:
x,hbnx,klnx,hlnx,kbn,hbn,kln,hln,sxbet,sxlet,sxbat, 8.1. The penultimate radical of the word obtained by
sxlat,sxbnx,sxlnx,sxbn,sxln,kubet,hubet,kulet,hulet trimming the suffix is geminated if the trimmed word
Three ,kubat,kulat,hulat,hubat,kubh,hulh,kubsx,hulsx,hub has three or more radicals, if its final radical is
Characters h,kulh,kulsx,hubsx,nbat,nlat,nbh,nlh,nbsx,nlsx,nlet first/second/fourth/sixth/unique order and its
,nbet,chuat,chut,kuacew penultimate radical is first order.
chuacew,cbacew,clacew,clachu,cbachu,kbacew,kla 8.2. The final radical of the word obtained by trimming the
cew,hbacew,hlacew,sxlacew,sxbacew,kulacew,kub suffix is geminated if the trimmed word is bi-radical and
Four acew,kubachu,hulachu,hubachu,kulachu,chubnx,ch its final radical is fourth order
Characters ulnx,chuln,chubn,chubat,chulat,chubet,chulet,nbac 9. If the verb has more than three radicals and has none of the
hu,nlachu,nbacew,nlacew suffixes listed in Table 2, the penultimate is geminated.
Five chubacew, chulacew
Characters Listing 2. Algorithm to assign geminations.
Let us use an example to show the one exception that is not
This scope is selected for three reasons: first, because it is almost handled by this algorithm. In the verb “Aጠራ” (atxrra) the
impossible to determine the gemination of other parts of speech consonant “ር” (r) is geminated; and in the verb “Aጠረ” (atxtxere)
just by analyzing the orthographic representation, it requires an the consonant “ጥ” (tx’) is geminated correctly according to rules

- 29 -
3 and 4, respectively. When both are inflected, “Aጠራቸው” “ስብክ”. The word “ሰበኩ” (sebbeku) has the suffix “U” (u)
(atxerrachew) or “Aጠራቸው” (atxtxerachew) may result. which makes the verb to be differently interpreted by supplying
However, the consonant “ር” (r) is geminated according to rule the subject “They” or “I”. Ignoring this ambiguity, the rules above
8.2, leaving out the possibility of geminating “ጥ” (tx’). This represent only the prosody model of the sentence having the
problem can only be resolved by using semantic disambiguation. subject “I”. This problem requires context disambiguation which
is out of the scope of this work.
3.3 Yes/No Question Prosody Representation
This section deals with modeling the interrogative and declarative 1. If a given statement ends with the symbol “?” then
prosodies. However, the modeling is focused on modeling the last 1.1. If its final word ends with one of the suffixes listed in
word of the sentences, and even from the last word, it is the last Table 2, select the interrogative sound unit version of
sound unit that is modeled. This is because the conspicuous the penultimate phoneme as the sound of the
difference between declarative and interrogative sentences is penultimate character in the word.
concentrated on the sentence-final word as the experiment from 1.2. But if its final word does not end with one of the
[5] concluded: “Judgment Requests were characterized by a suffixes listed in Table 2, select the interrogative sound
smaller standard deviation. This suggests that the distribution of unit version of the final phoneme as the sound of the
the amount of rise for Judgment Requests is relatively final character in the word.
concentrated and not scattered”. Yes/No interrogatives are further 2. Else if a given statement ends with the symbol “።”
divided into judgment and confirmation requests. In this work we 2.1. If its final word ends with one of the suffixes listed in
have focused on modeling judgment requests. We have performed Table 2, select the declarative sound unit version of the
waveform analysis to explore prosodic differences between penultimate phoneme as the sound of the penultimate
Amharic declarative and yes/no statements. The statements “Aበበ character in the word.
በሶ በላ።” (abebe beso bella.) and “Aበበ በሶ በላ?” (abebe beso 2.2. But if its final word does not end with one of the
bella?) are used here for comparison purpose. The figures below suffixes listed in Table 2, select the declarative sound
(Figure 2 and 3) show the difference of their waveforms. The unit version of the final phoneme as the sound of the
light/yellow thin lines show the amplitude (intensity) and the final character in the word.
black/blue bold lines on the spectrum show the pitch contour.
Listing 3. Algorithm to model sentence intonation.

3.4 Allophones
Each sound unit in the sound bank represents a phone or a syllable
based on the CV template only. Each non sixth order character in
the Amharic alphabet is represented by a minimum of three
allophones – one to represent phones that occur at the sentence
final (declarative’s falling pitch), another one to represent phones
that occur at the beginning or middle of a sentence, and still
another to represent the interrogatives’ rising pitch. In addition to
that, each of the first (ግEዝ), second (ካልE), fourth (ራብE) and
Figure 2. Waveform of the declarative statement.. unique (diphthongs) orders has one additional allophone to
represent the stressed version of that phoneme; and each of the
sixth orders (ሳድስ) also has one additional allophone to represent
the consonant phones without the epenthetic vowel ‘E’ (‘ɨ’).
Mark Beutnagel, et al. [11] proposes two methods to alleviate this
problem: specify greater allophonic detail in TTS front-end and
database labels, and identify contexts, such as pre-vocalic/post-
vocalic positions within a syllable, that determine, in part, the
allophonic variations. Our sound bank provides at least three
phones per each phoneme, to capture intonation differences that
surface because of gemination, epenthesis, punctuations (?,።) and
Figure 3. Waveform of the interrogative statement. their location in a sentence. The allophonic variations in the sound
bank are captured through a labeling scheme that facilitates easy
It is evident from Figure 2 and 3 that there is a conspicuous communication between the front and back end for fetching the
difference among the sentences’ final words and even further appropriate phone from the sound bank. However, since each non
analysis of the sentences’ final words reveals that the difference is sixth order alphabet has at least three phones associated with it,
concentrated on the final syllable of the word. It is clearly seen one extra character is appended to their Unicode to properly label.
that the amplitude (intensity) and the pitch contour of the The labeling (and representation) scheme looks like:
interrogative’s sentence-final word is greater than that of the [Unicode_value]: unstressed version of the phonemes
declarative. A set of rules (see Listing 3) is proposed for [Unicode_value]+‘0’: consonant phone of the 6th order characters
declarative and interrogative sentence modeling. The only [Unicode_value]+ ‘1’: geminated version of the phonemes
exception we have found for the rules given above is when the [Unicode_value]+ ‘2’: interrogative version of the phonemes
verb that comes at the end of the interrogative sentence ends with [Unicode_value]+ ‘3’: declarative version of the phonemes
the consonant “ክ” (k). One example can be given to clarify it;
when the word “ሰበኩ” (sebbeku) comes as inflection of the verb

- 30 -
4. EXPERIMENT by this approach are: having as many allophones as possible and
We developed an Amharic TTS system (implemented using C#) identifying contexts that determine allophonic variations. In this
that can synthesize virtually any Amharic text. All features– work, effort is shown to address these two factors. At least two
numeric transliteration, epenthesis, gemination, declarative versus allophonic variations for each phoneme in the Amharic alphabet
interrogative intonations are implemented. All the recordings are are built; and epenthesis, gemination and interrogative prosody
made by an individual who is a native speaker of the language modeling rules to make appropriate selection of those variations
with sampling rate of 48000 Hz. WaveSurfer™ 1.8.8pl- from context are also developed. We look forward to further
1010071431 and Praat™ 5.2.18 are the tools used for the research and extend the functionalities already covered in this
recording, labeling, waveform analysis and manipulation. system. Epenthesis rules that deal with geminated consonant
clusters, gemination rules that deal with Amharic words other than
4.1 Evaluation verbs in their perfect form and interrogative prosody modeling of
TTS systems require the evaluation of users by using techniques the WH-Questions are three of the features that we want to work
such as Mean Opinion Score (MOS) to measure their on as a sequel to this work.
intelligibility and naturalness. Ten individuals who are fluent
speakers of the language were selected to listen to sentences and 6. REFERENCES
words to evaluate their naturalness and intelligibility. The [1] Anne and Klaus Wedekind, “Amharic stress (beat) rules of
sentences and words were selected in such a way that they are linguists, poets and singers: Which beat rules beat which?”,
appropriate for evaluating major features of the synthesizer. As it The 11th Int. Conf. of Ethiopian Studies, Addis Ababa, 1991.
is described in the previous section, such features of the [2] Joyce McDonough, “The Prosody of Interrogative and Focus
synthesizer include: its ability to model prosody of a Yes/No Constructions in Navajo”, Department of Linguistics and
interrogative statements, declarative statements, properly assign Center for the Language Sciences, University of Rochester,
epenthesis and gemination. In the process, seventeen sentences Amsterdam, 2002
were prepared to evaluate whether prosodic features of
interrogative and declarative utterances were properly [3] Matthias Jilka and Ann Syrdal, “The AT&T German Text-to-
represented. Ten sentences are in declarative form, the remaining Speech System: Realistic Linguistic Description,” in
seven are in their interrogative form. Thirty words were selected Proceedings of ICSLP. 2002, Denver.
to evaluate the system’s control over gemination and epenthetic [4] Michael Gasser, “HornMorpho: a system for morphological
conditions. Ten Abbreviations abbreviated using the dot “.” analysis and generation of Amharic, Oromo, and Tigrinya
symbol and the forward slash symbol “/”and twelve numerals words”, Conference on Human Language Technology for
with and without decimal places were also included in the Development, Alexandria, Egypt, (2011).
evaluation. A questionnaire was prepared for the evaluation [5] Mieko Takada and Takako Ayusawa, “The Intonation of
process. The participants were made to hear the utterances without Interrogative Utterances in the Japanese Dialogs — Analysis
first looking at the texts; but after hearing they are shown the texts of the TUFS Language Module”. Yuji Kawaguchi, Iván
so that they can compare what they have heard against what they Fónagy, Tsunekazu Moriguchi, Prosody and Syntax: Cross-
have read and rank their opinion. linguistic Perspectives. Vol. 3, pp 239-263, 2006.
4.2 Result and Discussion [6] Mulugeta Seyoum, “The Syllable Structure and
The rating scales used for the MOS are: 5- very good, 4 – good, 3 Syllabification in Amharic,” MSc Thesis, Department of
– fair, 2 – poor, 1 – bad. The average score for the naturalness of Linguistics, Trondheim, Norway, 2001.
the system was 3.62 and its intelligibility 3.5 (see Table 3) which
[7] Nadew Tademe (2008). “Formant-based speech synthesis for
places the quality of the system just above fair and below good.
Amharic vowels”. MSc Thesis, Faculty of Informatics, Addis
The least scores are seen in epenthetic and numeral evaluations.
Ababa University, Ethiopia.
The low scores of epenthesis are related to the low quality of the
vowel-less consonants recorded whereas low scores on numerals [8] Ralf Kompe, “Prosody Speech Understanding Systems”,
is because of longer duration of pauses between numbers which Subseries of Lecture Notes in Computer Science. Springer-
fits well for non numeric text synthesis. Verlag, Berlin, 1997.
Table 3. MOS Results. [9] Sebsibe H/Mariam, S P Kishore, Alan W Black, Rohit
Kumar, and Rajeev Sangal. 2004. Unit Selection Voice for
Features Criteria Amharic Using Festvox, 5th ISCA Speech Synthesis
Naturalness Intelligibility Workshop, Pittsburgh.
Epenthesis 3.4 3.3
[10] Tadesse Anberber, Michael Gasser, Tomio Takara and Kim
Gemination 3.8 3.6
D. Yoon, “Grapheme-to-Phoneme Conversion for Amharic
Declarative Modeling 3.7 3.5 Text-to-Speech System”, Conference on Human Language
Interrogative Modeling 3.8 3.9 Technology for Development, Alexandria, Egypt, (2011).
Numerals 3.4 3.2 [11] Yeon-Jun Kim, Ann K. Syrdal, Alistair D. Conkie, Mark C.
Average 3.62 3.5 Beutnagel, “Phonetically Enriched Labeling in Unit
Selection TTS Synthesis,” INTERSPEECH, Pittsburgh,
5. CONCLUSION Pennsylvania, 2006.
One of the methods to easily produce intelligible and natural-like [12] Yibeltal Tefera (2008). “Formant-Based Speech Synthesis: A
synthetic sound is using the unit selection concatenative approach. Case of Amharic Words”. Addis Ababa University, Ethiopia.
Two factors that determine the quality of a TTS system developed

- 31 -