by
ESRA VURAL
Submitted to the Graduate School of Engineering and Natural Sciences
in partial fulllment of
the requirements for the degree of
Master of Science
Sabanc University
APPROVED BY KemalOFLAZER ... (Thesis Supervisor) Hakan ERDO GAN ... Yucel SAYGIN ... DATE OFAPPROVAL: ...18/07/2003...
SabancUniversity2003
IamthankfultomythesissupersivorKemalO azerforintroducingmetothis
chal-lengingsubject. Hisknowladge,patienceandinsight wasvery helpful inadvancing
through the project. I am alsograteful toHakan Erdogan for his motivations and
criticisms during the preparationof the thesis. I would also like to thank toYucel
Saygn forservingon mythesis committeeandforhis revisions onthe thesis.
I would like to thank to my parents and sister for their love, support and
en-couragement. Lastly,Iamthankfultothepatienceandsupportof AydnAkyolfor
Abstract
NaturalnessinText-to-Speechsystemsisveryimportantinachievinghigh
qual-ity waveform. The naturalness of the waveform is highlycorrelated with phonetic
coverage and prosodic features such as, duration and F0 contour. Duration
de-termines the timing for the synthesized phoneme, whereas F0 contour determines
fundamental frequency component of thewaveform.
This thesis presents the development of a prosodic Text-to-Speech System for
TurkishLanguageusingtheFestivalTool[31]. Wedescribeacompleterealizationof
anewmalevoice,coveringallophonesofTurkishusingdurationandF0parameters.
The duration of the allophonesand the word stress have been studied extensively.
Sentence stress andphrasal stress are alsodiscussed by inless detail.
Carrier words are designed approximately forall allophone-allophone
combina-tions. 1680 carrier words are recorded in a sound-proof recording studio. LPC
(linear predictive coding) and RES (residual) parameters are computed. The text
normalisationmoduleisimplementedforabbreviationsandnumbers. Durationsfor
the allophones are entered. Sentence level and word level F0 generation modules
are implemented. By increasing the number of phonemes and giving prosody we
TURKCE ICIN VURGULU METINDENSES SENTEZLEYICISI
Ozet
Metinden ses sentezleyicisi sistemlerinde dogallk kaliteli bir ses dalgas elde
edilmesinde cok onemli bir rol oynar. Ses dalgasnn dogallg fonetik kapsama ve
vurgusalozelliklerolanperdefrekansegrisive surebilgileriyleiliskilidir. Surebilgisi
sentezlenen fonemin zaman bilgisinibelirler, perde frekans egrisi ise ses dalgasnn
temelfrekans ozelliklerinikapsar.
Butezde,Festivalsessentezlemesistemikullanlarak,Turkceicinvurgulumetinden
ses sentezleyicisigelistirilmistir[31]. Yeni bir erkek sesi, Turkcedeki alofonlar
kap-sayarak, temel frekans ve sure bilgileri kullanlarak olusturulmustur. Alofonlarn
suresi ve kelime vurgusu genis capta cal slmstr. Cumle vurgusu ve kelime obek
vurgusu daha azdetaylolarakcalslmstr.
Tum alofon kombinasyonlar icin tasyc kelimeler olusturulmustur. 1680 tane
tasyc kelime ses yaltml bir kayt studyosunda kaydedilmistir. LPC ve RES
parametreleri hesaplanmstr. Ksaltmalar ve saylar icin metni normalize eden bir
modul gelistirilmistir. Alofonlar icin sure bilgisi girilmistir. Cumle ve kelime
se-viyelerinde F0 uretimmodullerigelistirilmistir. Fonem saysn arttrarak ve vurgu
Acknowledgments v
Abstract vi
Ozet vii
1 Introduction 1
1.1 Introduction ToSpeechSynthesis . . . 1
1.2 Text-to-Speech Systems . . . 1
1.3 Prosodic Turkish TTSSynthesizer. . . 1
1.4 Fundamental Dierences Between TTS Systems and Other Talking Machines. . . 2
1.5 Application AreasOfTTS Systems . . . 2
1.6 Review of PreviousWork . . . 3
1.6.1 CurrentWorkon TTSSystems . . . 3
1.6.2 TurkishTTSSystems . . . 5
2 Text-to-Speech 6 2.1 Stages OfTTS Conversion . . . 6
2.2 The NaturalLanguage ProcessingComponent . . . 7
2.2.1 TextAnalysis. . . 8
2.2.2 PhoneticAnalysis . . . 9
2.3 DigitalSignal ProcessingComponent . . . 11
2.3.1 ProsodicAnalysis . . . 11
2.3.2 Phonetics . . . 15
2.3.3 SpeechSynthesis . . . 15
3 The Festival Speech Synthesis System 19 3.1 Introduction ToFestivalSpeech Synthesis System . . . 19
3.2 Festival Text ToSpeech . . . 19
3.3 Utterance Structure. . . 19
3.4 Relations. . . 21
3.5 Modules . . . 22
3.6 Utterance Building . . . 24
3.8.1 ClusterUnitSelection . . . 29
3.8.2 Diphonesfrom generaldatabases. . . 30
3.9 Buildingprosodic models . . . 30
3.9.1 Phrasing . . . 30 3.9.2 Accent/BoundaryAssignment . . . 32 3.9.3 F0Generation . . . 33 3.9.4 F0byrule . . . 33 3.9.5 F0bylinearregression . . . 34 3.9.6 TiltModelling . . . 36 3.9.7 Duration . . . 36
4 A Prosodic Turkish Text-To-Speech System 39 4.1 TurkishPhonetization . . . 39
4.2 Stress inTurkish . . . 41
4.2.1 Roleof StressInTurkishWords . . . 41
4.2.2 PhoneticCorrelatesof Stress. . . 46
4.2.3 Distinctionsbetweendierentlevels ofstress . . . 46
4.2.4 Word-accent . . . 46
4.3 Sentence intonation . . . 48
4.4 DesigningandRecording ofa Diphone Corpus . . . 48
4.5 Text Normalization . . . 50
4.6 Designingthe Lexicon . . . 50
4.7 Designingthe Intonation . . . 53
4.8 Designingthe Duration . . . 57
5 Conclusion and Further Research 61 5.1 Conclusion . . . 61
5.2 FurtherResearch . . . 62
2.1 SimpleTTS Synthesis Procedure . . . 6
2.2 Basic SystemArchitecture of aTTS system . . . 7
2.3 Natural Language Processing module of a general TTS Conversion System . . . 8
2.4 DigitalsignalprocessingcomponentofageneralTTSconversionsystem 12 2.5 BlockDiagramof a prosodygeneration system . . . 13
2.6 Enriched ProsodyRepresentation . . . 14
2.7 Dierent kinds of informationprovided by intonation(lines indicate pitchmovements; solidlinesindicatestress[1]. a. Focusorgiven/new information;b. Relationshipsbetweenwords(saw-yesterday;I-yesterday; I-him) c. Finality (top) or continuation (bottom), as it appears on the lastsyllable; . . . 14
2.8 The human vocalorgans . . . 16
2.9 Blockdiagramof a synthesis-by-rulesystem . . . 16
3.1 An example representation of an utterance structure. This example showsthewordrelationandthesyntax relation. Thesyntax relation (shown on top) is a tree with links connecting the nodes, shown as blackcircles. Thewordrelation (shownon thebottom)isalist. The items containthe actuallinguisticinformationandare shown in the rounded boxes. The dotted lines show the connections between the nodes anditems. . . 21
3.2 Close-up pitchmarks inwaveform signal . . . 28
3.3 TOBI Parameters . . . 35
2.1 UnittypesinEnglish assuming aphone set of 42phonemes. Longer
Units produce higherqualityattheexpense of morestorage. . . 17
4.1 TurkishVowel Inventory . . . 40
4.2 TURKISH PHONETICENCODING FOR VOWELS . . . 41
4.3 TURKISH PHONETICENCODING FOR VOWELSCONTINUED 42 4.4 TURKISH PHONETICENCODING FOR CONSONANTS . . . 43
4.5 ALLOPHONES USED IN TTSFOR VOWELS . . . 44
4.6 ALLOPHONES USED IN TTSFOR CONSONANTS . . . 45
4.7 Letter toSound Conversion Table . . . 52
4.8 Duration ofthe allaphonesinmilliseconds[36].. . . 59
by
ESRA VURAL
Submitted to the Graduate School of Engineering and Natural Sciences
in partial fulllment of
the requirements for the degree of
Master of Science
Sabanc University
APPROVEDBY Kemal OFLAZER ... (Thesis Supervisor) HakanERDO GAN ... Yucel SAYGIN ...
Iamthankful tomythesissupersivorKemalO azerforintroducingmetothis
chal-lengingsubject. His knowladge, patience and insightwas very helpful in advancing
through the project. I am alsograteful to Hakan Erdogan for his motivations and
criticismsduring the preparation of the thesis. I would alsolike tothank to Yucel
Saygnfor serving on my thesis committee and for hisrevisions on the thesis.
I would like to thank to my parents and sister for their love, support and
en-couragement. Lastly,I amthankfultothe patience and supportof AydnAkyolfor
Abstract
Naturalness inText-to-Speechsystems isvery importantinachievinghigh
qual-ity waveform. The naturalness of the waveform is highlycorrelated with phonetic
coverage and prosodic features such as, duration and F0 contour. Duration
de-termines the timing for the synthesized phoneme, whereas F0 contour determines
fundamentalfrequency component of the waveform.
This thesis presents the development of a prosodic Text-to-Speech System for
TurkishLanguageusingtheFestivalTool[31]. Wedescribeacompleterealizationof
anewmalevoice,coveringallophonesofTurkishusing durationandF0parameters.
The duration of the allophones and the word stress have been studied extensively.
Sentence stress and phrasal stress are alsodiscussed by inless detail.
Carrier words are designed approximately for all allophone-allophone
combina-tions. 1680 carrier words are recorded in a sound-proof recording studio. LPC
(linear predictive coding) and RES (residual) parameters are computed. The text
normalisationmoduleisimplementedforabbreviations andnumbers. Durationsfor
the allophones are entered. Sentence level and word level F0 generation modules
are implemented. By increasing the number of phonemes and giving prosody we
TURKCE ICIN VURGULU METINDEN SES SENTEZLEYICISI
Ozet
Metinden ses sentezleyicisi sistemlerinde dogallk kaliteli bir ses dalgas elde
edilmesinde cok onemli bir rol oynar. Ses dalgasnn dogallg fonetik kapsama ve
vurgusalozelliklerolanperdefrekansegrisivesurebilgileriyleiliskilidir. Surebilgisi
sentezlenen fonemin zaman bilgisini belirler, perde frekans egrisi ise ses dalgasnn
temel frekans ozelliklerinikapsar.
Butezde,Festivalsessentezlemesistemikullanlarak,Turkceicinvurgulumetinden
ses sentezleyicisi gelistirilmistir[31]. Yenibirerkek sesi, Turkcedeki alofonlar
kap-sayarak, temel frekans ve sure bilgileri kullanlarak olusturulmustur. Alofonlarn
suresi ve kelime vurgusu genis capta cal slmstr. Cumle vurgusu ve kelime obek
vurgusu daha azdetayl olarakcalslmstr.
Tum alofon kombinasyonlar icin tasyc kelimeler olusturulmustur. 1680 tane
tasyc kelime ses yaltml bir kayt studyosunda kaydedilmistir. LPC ve RES
parametreleri hesaplanmstr. Ksaltmalar ve saylar icinmetni normalizeeden bir
modul gelistirilmistir. Alofonlar icin sure bilgisi girilmistir. Cumle ve kelime
se-viyelerinde F0 uretim modulleri gelistirilmistir. Fonem saysnarttrarak ve vurgu
Acknowledgments v
Abstract vi
Ozet vii
1 Introduction 1
1.1 Introduction To Speech Synthesis . . . 1
1.2 Text-to-Speech Systems . . . 1
1.3 Prosodic Turkish TTS Synthesizer . . . 1
1.4 Fundamental Dierences Between TTS Systems and Other Talking Machines. . . 2
1.5 ApplicationAreas Of TTS Systems . . . 2
1.6 Review of Previous Work . . . 3
1.6.1 CurrentWorkon TTSSystems . . . 3
1.6.2 TurkishTTSSystems . . . 5
2 Text-to-Speech 6 2.1 StagesOf TTS Conversion . . . 6
2.2 The Natural Language Processing Component . . . 7
2.2.1 Text Analysis. . . 8
2.2.2 PhoneticAnalysis . . . 9
2.3 DigitalSignal ProcessingComponent . . . 11
2.3.1 ProsodicAnalysis . . . 11
2.3.2 Phonetics . . . 15
2.3.3 SpeechSynthesis . . . 15
3 The Festival Speech Synthesis System 19 3.1 Introduction To Festival Speech Synthesis System . . . 19
3.2 FestivalText To Speech . . . 19
3.3 Utterance Structure . . . 19
3.4 Relations. . . 21
3.5 Modules . . . 22
3.6 Utterance Building . . . 24
3.7 Diphone Databases . . . 25
3.8.1 ClusterUnit Selection . . . 29
3.8.2 Diphones fromgeneral databases. . . 30
3.9 Buildingprosodic models . . . 30
3.9.1 Phrasing . . . 30 3.9.2 Accent/Boundary Assignment . . . 32 3.9.3 F0Generation . . . 33 3.9.4 F0byrule . . . 33 3.9.5 F0bylinearregression . . . 34 3.9.6 TiltModelling . . . 36 3.9.7 Duration . . . 36
4 A Prosodic Turkish Text-To-Speech System 39 4.1 Turkish Phonetization . . . 39
4.2 Stress inTurkish . . . 41
4.2.1 Roleof Stress InTurkishWords . . . 41
4.2.2 PhoneticCorrelates ofStress . . . 46
4.2.3 Distinctionsbetweendierentlevels of stress . . . 46
4.2.4 Word-accent . . . 46
4.3 Sentence intonation . . . 48
4.4 Designing and Recording of aDiphone Corpus . . . 48
4.5 Text Normalization . . . 50
4.6 Designing the Lexicon . . . 50
4.7 Designing the Intonation . . . 53
4.8 Designing the Duration . . . 57
5 Conclusion and Further Research 61 5.1 Conclusion . . . 61
5.2 FurtherResearch . . . 62
2.1 SimpleTTS Synthesis Procedure . . . 6
2.2 BasicSystem Architecture of aTTS system . . . 7
2.3 Natural Language Processing module of a general TTS Conversion System . . . 8
2.4 DigitalsignalprocessingcomponentofageneralTTSconversionsystem 12 2.5 Block Diagramof aprosody generation system . . . 13
2.6 Enriched Prosody Representation . . . 14
2.7 Dierent kinds of information provided by intonation (lines indicate pitchmovements;solidlinesindicatestress[1]. a. Focusorgiven/new information;b. Relationshipsbetweenwords(saw-yesterday;I-yesterday; I-him) c. Finality (top) or continuation (bottom), as it appears on the lastsyllable; . . . 14
2.8 The human vocalorgans . . . 16
2.9 Block diagramof a synthesis-by-rule system . . . 16
3.1 An example representation of an utterance structure. This example shows theword relationandthe syntax relation. Thesyntax relation (shown on top) is a tree with links connecting the nodes, shown as blackcircles. The word relation(shown onthe bottom)isalist. The items contain the actual linguistic informationand are shown in the rounded boxes. The dotted lines show the connections between the nodes and items. . . 21
3.2 Close-up pitchmarks inwaveform signal . . . 28
3.3 TOBI Parameters . . . 35
3.4 TiltParameters . . . 37
2.1 Unittypes inEnglish assuming aphone set of 42phonemes. Longer
Unitsproduce higher quality atthe expense of more storage. . . 17
4.1 Turkish Vowel Inventory . . . 40
4.2 TURKISHPHONETIC ENCODING FORVOWELS . . . 41
4.3 TURKISHPHONETIC ENCODING FORVOWELSCONTINUED 42
4.4 TURKISHPHONETIC ENCODING FORCONSONANTS . . . 43
4.5 ALLOPHONES USED IN TTS FOR VOWELS . . . 44
4.6 ALLOPHONES USED IN TTS FOR CONSONANTS . . . 45
4.7 Letter toSound Conversion Table . . . 52
4.8 Duration of the allaphones inmilliseconds [36]. . . 59
Introduction
1.1 Introduction To Speech Synthesis
Speech is the one of the most eective means of communication between people.
Synthesizing is the process of generating speech waveforms using machines based
onthe phonetical transcriptionof themessage. Recent progress inspeech synthesis
has produced synthesizers with very high intelligibility but the sound quality and
naturalness stillremain amajorproblem.
1.2 Text-to-Speech Systems
AText-to-Speech(TTS)Systemisasystem thatconvertsany formofcomputerized
text into speech waveforms [1]. In other words, a Text-to-Speech system emulates
a real human speaker that can read any text aloud.
1.3 Prosodic Turkish TTS Synthesizer
This thesis presents the implementation of a prosodic Turkish TTS Synthesizer.
The prosody is generated by dening durationand intonationmodules considering
the language specic properties of Turkish. The format of a Turkish Lexicon is
described for the transcription of ortographic form into phonetic representation.
ing Machines
The fundamental dierence between TTS systems and other talking machines like
casette-players lies under the fact that TTS systems generate all the words
syn-thetically whereas other talking machines may concatenate pre-recorded words or
sentencesand generatespeech. Inthis respect, TTSsystems maygenerate anytext
inputintospeechwaveformeasily. SoTTSsystemsareveryimportantintasksthat
requirelargevocabulary. TTSsystemsuse graphemetophonemeconversionforthe
automatic generation of speech.
1.5 Application Areas Of TTS Systems
In the mid-eighties, as a result of the improved techniques in natural language
processing and signal processing, the concept of high quality TTS appeared. TTS
systems became useful in many commercialand personalusage.
Here are some examplesof high qualityTTS applications:
TelecommunicationServices: TTS systems makeiteasytoaccess textual
information over the telephone. Texts may range from simple messages
con-sistingfromasmallcorpus,tohugecomplexmessagesthatcanneverbestored
as speech waveforms. Queries to information retrieval systems can be made
throughspeech(with thehelpof aspeechrecognizer)and theanswer
(synthe-sized text) isreturned back tothe caller. AT&T has developed TTS systems
for applicationssuch asintegrated messaging and telephone relay service.
1)IntegratedMessaging(electronicmailorfascimilereaderapplication),which
is avery useful application when one is away from the oce.
2)TelephoneRelay Service,anapplicationthatisused forhavingatelephone
conversation with a hearing or speech impaired person. These applications
were acceptable althoughthe sythesized voice wasnot sounding very natural.
LanguageEducation: Highqualitysynthesiscanbecombinedtogetherwith
lectures and presentations using TTS technology. A special keyboard that is
designed forhandicapped peopleand asynthesizercan becoupledtohelp the
voice handicapped. Blind people may also use TTS systems when coupled
withanOCR(OpticalCharacterRecognition)System, thiscanmakepossible
access toany written text.
Talking Books and Toys: Talking books and toys are not new to market,
butthe qualityof thesynthesis becomesmoreimportantinthe entertainment
business.
Vocal Monitoring: In some situations, oral information may be more
ap-propriatethanwritteninformation. TTSsystemscanbeusedinmeasurement
and controlsystems.
Multimedia and Man-Machine Communication: The interaction
be-tween man and the machines willbeindispensablein the near future so TTS
systems must beable to produce highlyqualied speech synthesis.
Fundamental and Applied Research: TTS systemsare excellenttoolsfor
linguistic research since they have constant, stable features, in other words
repeated experience provides identical results. Even the people can change
theirutterances. SoTTS systemsare excellent laboratorytoolsfor intonative
and rhythmic models.
1.6 Review of Previous Work
1.6.1 Current Work on TTS Systems
There are many TTS systems for various languages. In the past synthesizers were
built only considering a single language. Thus converting this language specic
synthesizer for another language was a tough job. Reinventing the wheel for every
newlangugewasawasteof timeand eort. Instead,researchlabs decided toinvest
their eort in developing multilingual synthesizers. Their aim is developing TTS
synthesizers for as many languages, dialects and voices as possible. In this section,
nique de Mons, Belgium. The aim of this project is todevelop multilingualspeech
synthesisfornon-commercialpurposesandincreasetheacademicresearch,especially
in prosody generation. MBROLA method is similar to PSOLA method. But it is
named MBROLA since PSOLA is a trademarkof CNET. The MBROLA-material
is available free for non-commercial and non-military purposes [2]. The MBROLA
synthesizerisbasedondiphoneconcatenation. Ittakesalistofphonemeswithsome
prosodicinformationnamely durationand pitch asinput. The input data required
by MBROLA contains aphoneme name,a durationin milliseconds,and a series of
pitch pattern points which are two integers. For instance, the input " 80 30 130"
describes that the synthesizer willproduce a silence of 80ms, and willput a pitch
pattern points of 130 Hz at 30% of 80 ms. MBROLA produces speech waveforms
of16bits atthe samplingfrequency of16kHz. It isnotaTTS system sinceitdoes
not accept raw text as input but it may be used as a low level synthesizer. The
diphone databases are currently available for American/British English, Brazilian
Portugese, Dutch,French, German,Romanian,Spanishand Turkish withmale and
femalevoices.
FestivalTTS system wasdeveloped in CSTRatthe University ofEdinburgh by
AlanBlackand PaulTaylor. The basicarchitecture of Festivalwasin uenced from
ATR'sCHATRsystem. ThesystemiswritteninC++. FestivalTTSSynthesis
sys-tem isavailable forAmericanand BritishEnglish,Spanishand Welsh. This system
supports residualexcited LPC and PSOLA methodsand MBROLA database. The
system is availablefree for educational, research and individual use. The system is
developed for three main user groups, those who want to use the system for
arbi-traryTTS,forthosewho aredeveloping languagesystems andnallyforthosewho
are developing new synthesis methods.
TheUniversitedeProvenceiscoordinatoroftheMULTEXT[3]seriesofprojects.
The aim is developing tools and corpora and linguistic resources for a variety of
Bozkurt [4] describes a synthesizer for Turkish using the MBROLA technique.
MBROLA technique attempts to overcome phase mismatches. As the pitch
cy-cles have been pre-processed they have a xed phase. It uses the PSOLA method
formodifyingtheprosody. Spectralsmoothingcanbedonebydirectlyinterpolating
the pitch cycles inthe time domain.
Another TTS project for Turkish has been developed by Levent Arslan from
Bosphorous University and his team. They used LPC modelling for the
parame-terization of the speech waveform. This TTS System assumes that the context of
the current phoneme is eected from the most closest two phonemes both on the
left and right. The synthesis module searches for the most appropriate lpc le in
thedatabase, extractsLPCparametersfromthatleandusesduration, F0contour
with these parameters to synthesize frames. Durations are determined using two
types of information. Half of the duration information comes from the model and
the other half comes from actualwaveforms. Prosody module nds the most likely
pitchcontour ofsynthesis. Forvoicedregions,theprosodymodulecalculatesa
non-zeroperiodvaluewhichisusedingeneratinganimpulsetrain. Forunvoicedregions
Text-to-Speech
In this chapter, we willdescribe the phases of a Text-to-Speech system. The
com-ponentsthat will be implementedfor a TTS system willbeexplained indetail.
2.1 Stages Of TTS Conversion
TheTTSsynthesisprocessconsistsoftwomainphasesasshown inFigure 2.1. The
Text and linguistic
analysis
Input Text
Phonetic
Level
Prosody and Speech
Generation
Synthesized
Speech
Figure 2.1: Simple TTSSynthesis Procedure
rst phase is text analysis, where the input text is transcribed into some
linguis-tic and phonetic representation. This phonetic representation usually includes the
phonemes,duration, stressandintonation. The secondphase isprosodyandspeech
generation where the linguistic information is used to produce the correct speech
waveform. The rst part isimplemented by a NaturalLanguage Processing(NLP)
module that is capableof producinga phonetic transcription of the given text,
to-gether with the intonation. The second part is implemented by a Digital Signal
Processing (DSP) module, that transforms the phonetic and prosodic information
into speech.
The basic TTS components are shown in Figure 2.2 [37]. The text analysis
pro-nally, the speech synthesis component takes the parameters from the fully tagged
phonetic sequence to generate the corresponding waveform. These modules willbe
described indetail later.
Raw Text
Text Analysis
Document Structure Detection
Text Normalisation
Linguistic Analysis
Phonetic Analysis
Grapheme-to-Phoneme Conversion
Tagged Text
Prosodic Analysis
Pitch & Duration Assignment
Speech Synthesis
Voice Rendering
Text-To-Speech Engine
Tagged Phonemes
Controls
Figure2.2: Basic System Architecture of a TTS system
2.2 The Natural Language Processing Component
Figure 2.3 [37] shows the skeleton of a general natural language processing
mod-ule for TTS systems. The module consists of text analysis and phonetic analysis
Document Structure Detection
Text Normalisation
Linguistic Analysis
Homograph Disambiguation
Morphological Analysis
Letter-to-Sound Conversion
Tagged Text & Phones
Tagged Text
Raw Text
Document Structure Detection
Text Normalisation
Linguistic Analysis
Homograph Disambiguation
Morphological Analysis
Letter-to-Sound Conversion
Tagged Text
Raw Text
Text
Analysis
Phonetic
Analysis
Lexicon
Figure2.3: NaturalLanguage Processingmoduleof ageneralTTSConversion
Sys-tem
2.2.1 Text Analysis
The text analysis component is responsible for determining document structure,
conversion ofnon-ortographic symbols,and identicationoflanguagestructure and
meaning. Simple TTS systems justconvert the nonortographic items such as
num-bers, into words. More ambitious TTS systems attempt to analyze white spaces
andpunctuationsinorder todetermine both documentstructure andsyntactic and
semantic structure of the text. The syntactic and semantic structure enables the
system to determinethe phonetic and prosodicrepresentation. The modules of the
text analysis componentwill bedescribed below.
i.e sentence starts, paragraph starts. Document Structure eects the prosody. For
instance, thebeginningof asentence must bedetectedinordertoattachadierent
prosodicpattern. Likewise,aparagraphstart shouldbeidentiedsothat asuitable
prosodicpatterncan begiven.
Text NormalisationModule:
Text normalisation is the conversion of abbreviations, numbers, symbols and
othernon-ortographic entitiesof textintoacommonortographictranscriptionthat
is suitable for subsequent phonetic conversion. It is the process of generating the
ortography of the given text. In the following examplenumber and percent sign is
transformed into writtenform [37].
The 5% mixture -> THE FIVE PERCENT MIXTURE
Linguistic AnalysisModule:
Linguistic Analysis Module determines the syntactic and semantic features of
words, phrases, clauses and sentences. This module helps to resolve
grammati-cal features and part of speech of individual words. Parsing also helps for
deriv-ingprosodicstructure useful indetermining segmentalduration and pitchcontour.
Parsing can contribute tothe syntactic typeof the sentence. The prosody depends
on the question type of the sentence, yes/no question sentences will dier in their
prosodycontour from who/wheresentences.
2.2.2 Phonetic Analysis
The phonetic analysismoduleisresponsible forconvertinglexicalortographic
sym-bols tophonetic representation with stress information. The following modules are
needed to haveaccurate pronounciations.
Homograph Disambiguation:
Sense disambiguity occurs when words have dierent syntactic/semantic
mean-ings. Homograph disambiguation module disambiguates these ambiguities. Words
with dierent senses are called polysemous words. Polysemous words that are
pro-nounced dierently are called homographs. Some homographs such as object,
ab-sent,minuteetc canberesolved bytheirpart-of-speech. Objectispronouncedas(/
the context.
Morphological Analysis:
The morphological analysis module relates a surface ortographic form to its
pronounciation by analyzing its component morphemes, such as prexes, suxes
and stem words. This decomposition process is referred as morphological analysis
[12].
Letter-to-sound Conversion:
Letter-to-sound conversion is the last stage of phonetic analysis where general
letter-to-soundrulesareappliedandadictionarylookupismadetoproduceaccurate
pronounciationsforanarbitraryword. Theletter-to-sound (LTS) module
automat-icallydeterminesthephonetic transcriptionof theincomingtext. Themost reliable
way for grapheme to phoneme conversion is via dictionary lookup. If dictionary
lookup fails, rules may be used to generate the phonetic forms. An example rule
can be given asfollows:
ortographic k can be changed to a velar 1 plosive 2 /k/ when k is word initial('[') followed by n can be given as k -> /sil/ % [ _ n k -> /k/
The rewrite rules above indicate that k is transformed into silence when it is in
word initialposition and followed by n, otherwise it is rewritten as phonetic /k/.
The underscore inthe rst lineis a placeholder for the k itself. The word knight is
an example for the realisationof the letter k into silence. Generallya TTS system
requires hundreds of these rules for converting words that are not present in the
exception list, or they may be listed oine in a lexicon. We can organize the LTS
module's taskintotwo dierentways, dictionarybased and rule based:
1
Velarconsonantsbringthebackofthetongue,uptotherearmosttopareaoftheoralcavity
2
a lexicon. In order to keep its size small, entries are restricted to morphemes.
The pronounciation of the surface forms is made by in ectional, derivational and
compoundingmorphophonemicruleswhichdescribehowthephonetictranscriptions
oftheir morphemicconstitutents arecombinedintowords. Morphemes that cannot
befoundinthelexiconaretranscribedbyrules. Afterarstphonemictranscription
ofeachword has been obtained,somephonetic post-processingisgenerallyapplied,
so as to account for coarticulatory smoothing phenomena [1]. This approach has
been appliedby theMITTALKsystem [9]. Adictionaryofup to12,000morphemes
coveredabout 95%oftheinputwords. TheAT&TBellLaboratoriesTTSsystemis
quitesimilar [10], with anaugmented morpheme lexicon of 43,000morphemes [11].
O azer and Inkelas [14] have implemented a full scale pronounciation lexicon for
Turkish using nite state technology. The system outputs a word's morphological
and pronounciation forms.
Rule based solutions transcribe most of the words into phonological forms by
applying the phonological rules. Onlywords that are pronounced in one particular
way are stored in an exception dictionary. Notice that, since many exceptions are
found in the most frequent words, a reasonably small exceptions dictionary can
accountfor alarge fractionof the words inarunningtext. In English,for instance,
2000 words typicallysuce tocover 70% of the words in text[22].
2.3 Digital Signal Processing Component
Figure 2.4 shows the skeleton of a general digital signal processing module for
TTS systems [37]. The module consists of Prosodic Analysis and Speech Synthesis
components:
2.3.1 Prosodic Analysis
Prosodicfeatures have specicfunctionsinspeechcommunication. Mostimportant
function of the prosodic features is to focus a single or a group of words. For
in-stance, the pitch of a syllable may stress a certain word and this may change the
wholemeaningof the utterance. In otherwords, certainwords must be highlighted
Document Structure Detection
Text Normalisation
Linguistic Analysis
Homograph Disambiguation
Morphological Analysis
Letter-to-Sound Conversion
Tagged Text & Phonemes
Tagged Text
Raw Text
Document Structure Detection
Text Normalisation
Linguistic Analysis
Homograph Disambiguation
Morphological Analysis
Letter-to-Sound Conversion
Tagged Text
Raw Text
Text
Analysis
Phonetic
Analysis
Lexicon
Figure2.4: DigitalsignalprocessingcomponentofageneralTTSconversionsystem
properties of the speech signal which are related to audiblechanges inpitch,
loud-ness, syllable length [1]. It expresses linguistic information such as sentence type
and phrasing as well as paralinguistic 3
information such as emotion. It is
charac-terized by three main parameters: fundamental frequency, duration and intensity.
Figure 2.5 shows the components of the prosodic analysis module [37]. Enriched
prosodyrepresentation(asseeninFigure 2.6)consistsoflinesofinformationwhere
eachline speciesthe phoneme name,the phoneme durationin milliseconds,and a
numberofprosodypointsspecifyingpitchandsometimesvolume[37]. Forinstance,
the fourth linespecies values forphoneme /k/whichlasts 80ms. Italsohas three
prosodictargets. The rst is located at 25% of the phoneme duration, i.e., placed
atthe 20ms of the phoneme durationand has apitch value of 178 Hz.
The durationof segments is dicultto describe. It correlates with many
prag-matic categories, like sentence focus, emphasis and stress [13]. One can synthesize
each phoneme with its average duration. Alternatively, duration can be specied
by hand-writtenrules [15] ornumericalmodels [16]. Although durationcan be
de-scribed categorically,itis essentially acontinuous variable,and should thereforebe
well-suitedtomethodslikenon-linearregression [16],regression trees[17], orneural
3
Pitch-range variationthat is correlated withemotion orother aspects ofthe speech eventis
Pause Insertion and Prosodic Phrasing
Duration
F0 Contour
Volume
Pause Insertion and Prosodic Phrasing
Duration
F0 Contour
Volume
Parsed text and Phone String
Enriched Prosodic Representation
Figure 2.5: Block Diagramof aprosody generation system
networks. However, statistical models or rule-based models require large amounts
of labelled training data, which are dicult to produce. Therefore, using average
phoneme durations appears to be the onlyviablesolution.
Intonationisrepresentedintwodierentways. Oneisbasedonthepitchcontour
of an utterance. Pitch contour is a sequence of rises and falls of F0 [18]. The
other way of representation [19]describesintonationas asequence of pitch targets.
Thereare twolevelsoftargetsortones: high(H,maximum)and low(L,minimum).
These targets mark either a pitch accent, that is, extremum in the pitch contour,
ora prosodicboundary (boundary tone). The target that bears anuclear accentis
starred,boundarytonesaremarkedbythesign%. TOBI(TonesandBreakIndices)
[21] is one of the most popular phonological labelling system, which is used widely
for annotating prosodiccorpora. In TOBI (Tones and Break Indices) labelling, the
tones are coupled with break indices[23] for markingprosodicphrase boundaries.
Bothapproachesare onaphonologicallevel: theyareused todescribetherough
pitchcontourofanutteranceandhavetobetransformedintophoneticdescriptions.
If one wants tosynthesize a pitchcontour from a TOBI (Tones and Break Indices)
labelling,oneneedstoknowtowhatvaluesthe tonescorrespond, howtointerpolate
DH, 24
(0,178);
AH0, 104
;
# ;
K, 80 (25,178)
(50,184)
(75,201);
AE1, 152 (0,214)
(25,213)
(50,204)
(75,193);
T, 40 (0,175)
(25,175)
(50,174)
(75,172);
# ;
K, 104 (0,171)
(25,172)
(50,180)
(75,189);
AE1, 104 (0,198)
(25,196)
(50,168)
(75,137);
T, 112 (0,120)
(100,200);
# ;
Figure2.6: Enriched Prosody Representation
I saw him yesterday
I saw him yesterday
I saw him yesterday
I saw him yesterday
I saw him yesterday
I saw him yesterday
I saw him yesterday
I saw him yesterday
a.
b.
c.
Figure 2.7: Dierent kinds of information provided by intonation (lines indicate
pitchmovements; solid lines indicatestress [1]. a. Focus orgiven/new information;
b. Relationshipsbetweenwords(saw-yesterday; I-yesterday;I-him)c. Finality(top)
orcontinuation (bottom),as itappears on the lastsyllable;
are blocked by those boundaries. Furthermore, we need to set global parameters
for the pitch contour likea top line and a base line, which bound pitch excursions.
During speaking, fundamental frequency declines slowly, and this fact has to be
considered too.
Stress is an important information carrier. Word stress determines which
syl-lable in a word is stressed, phrase stress determines the words in a phrase that
receive stress. Stress can be signalled by all prosodic parameters pitch, intensity,
and duration, as well as by segment quality. Stressed syllables tend to be longer
than unstressed ones, and they are usually further marked by alocalmaximum or
Human speech is produced by vocal organs as presented in Figure 2.8. The main
energysourceisthelungswiththediaphragm. Whenspeaking,theair owisforced
throughtheglottisbetweenthevocalcordsandthelarynxtothethreemaincavities
of the vocal tract, the pharynx and the oral and nasal cavities. From the oral and
nasal cavities the air ow exits through the nose and mouth, respectively. The
V-shaped opening between the vocal cords, called the glottis, is the most important
sound source in the vocal system. The most important function of vocal cords is
to modulate the air ow by rapidly opening and closing, causing buzzing sound
fromwhichvowelsandvoicedconsonantsareproduced. Thefrequency ofvocalfold
vibration is called fundamental frequency. The fundamentalfrequency of vibration
dependsongenderandphysicalpropertiesofthemouthandnosecavityandisabout
110 Hz, 200 Hz, and 300 Hz with men, women, and children, respectively. With
stop consonants the vocal cords act suddenly from a completely closed position in
which they cut the air ow completely, to totally open position producing a light
cough or aglottal stop. On the other hand, with unvoiced consonants, such as /s/
or/f/,theymaybecompletelyopen. Anintermediatepositionmayalsooccurwith
for example phonemes like /h/.
2.3.3 Speech Synthesis
Intuitively, the operations involved in the digital signal processing module are the
computeranalogueofdynamicallycontrollingthearticulatorymusclesandvibratory
frequencyofthevocalfoldssothattheoutputsignalmatchestheinputrequirements
[1]. Ithas beenknown fora longtimethat phonetic transitionsare moreimportant
thanstable statesforthe understanding ofspeech[24]. This canbeachieved intwo
ways:
Rule Based/Formant Synthesizers:
Rule-based synthesizers havetwo major components: the rst is agenerator for
the excitation signaland the secondis alter that simulates the eect of the vocal
tract. The parametersofthe lterare derived fromacousticspecications. Inother
I saw him yesterday
I saw him yesterday
Figure2.8: The human vocal organs
lter with the formantfrequencies of the vocal tract. When synthesizing unvoiced
speechwhiterandomnoise 4
can beusedforthesourceinstead. Sincespeechsignals
arenotstationarythepitchoftheglottalsourceandtheformantfrequencieschange
overtime. Synthesis-by-rule refers toaset of rulesonhowtomodifypitch,formant
frequencies and other parametersfrom one soundto anotherwhile maintainingthe
continuity present in physical systems like the human production system. Figure
2.9 is a model of such a system. Rule-based synthesizers provide great assistance
Phonemes +
Prosodic Tags
Rule-based
System
Pitch Contour
Formant
Track
Formant
Synthesizer
Output
Waveform
Figure 2.9: Block diagramof a synthesis-by-rule system
nism. The spread of the Klatt synthesizer [6] is due to its invaluable assistance in
the study of the characteristicsof naturalspeech. The relationships between
artic-ulatory parameters and the inputs of the Klatt model make it a practical tool for
investigatingphysiologicalconstraints [25].
ConcatenativeSynthesizers:
Concatenative speech synthesis systems generate speech by concatenating and
manipulating prerecorded units of speech. Choice and storage of units and
con-catenation method are important in this method. Coarticulatory eects have to
be modelled using a potentially limited unit inventory. A balancehas to be found
between the quality and the size of the inventory. As the units get larger, quality
increases but onthe other hand more units have toberecorded and segmented. A
compilation of unit types for English isshown inTable 2.1.
Unit Length Unit Type Number of Units Quality
Short Phoneme 42 Low
Diphone 1500 Triphone 30K Demisyllable 2000 Syllable 15K Word 100K-1.5M Phrase 1
Long Sentence 1 High
Table 2.1: Unit types in English assuming a phone set of 42 phonemes. Longer
Unitsproduce higher quality atthe expense of more storage.
The rst approach used in concatenative synthesis was concatenating single
phonemes [26]. Having one instance of each phoneme, independent of the
neig-bouringcontext,isverygeneralizable. Itallowsustogenerate everyword/sentence.
Thismethodyieldedratherpoorquality sincecontext-independentphonesresult in
many audiblediscontinuities.
the word hello /hh ax l ow/, we have to concatenate the diphones /sil-hh/ ,
/hh-ax/ , /ax-l/ , /l-ow/ , and /ow-sil/. While using the diphone concatenation, we
must assume that the transition between the two phones is sucient to model all
neccessary coarticulatory eects. The other assumption is that the spectra of the
steady states of the phones are consistent enoughto avoidspectral discontinuties.
Comparison of Rule-Based and Concatenative Synthesis:
Withrulebasedsystems,onecanadjustahighnumberofparametersandobtain
high-quality speech. Discontinuties arising fromconcatenating pre-recorded speech
units is not a problem with this method. However, setting parameters and
devis-ing rule sets such that the resulting speech is both intelligible and natural is very
dicult.
On the other hand, aconcatenative approach onlyneeds the phonetic layout of
the language, a good concatenation algorithmand apatientspeaker. A
concatena-tive synthesizer is also more natural when compared with rule based synthesizers,
The Festival Speech Synthesis System
Weusedthe Festival[31]system forthe work presented inthis thesis. This chapter
describes the relevantdetails of this system.
3.1 Introduction To Festival Speech Synthesis System
Festival oers a general framework for building speech synthesis systems and also
includesexamplesofvariousmodules. ItprovidesuswithafullTTSsystemthrough
API's 1
: using Scheme command interpreter, or C++ library. It is a multi-lingual
system. Voices in English (UKand US), Spanish, Welsh have been developed with
this system. The system is writtenin C++ and mainlyuses the EdinburghSpeech
Tools for low level architecture and has a Scheme based command interpreter for
control.
3.2 Festival Text To Speech
Festivalsupports text-to-speech for rawtext les. Atcommand line,
festival -tts textfile
renders input textleintospeech waveform. In other words this command willsay
the contents ofthe textle.
3.3 Utterance Structure
The basic building block for Festivalis the utterance. Utterance structure consists
of a set of relations over a set of items. Items represent objects such as words,
1
Relationstakeany graphstructure but they are commonlylistsor trees. Relations
relatetheitems. Itemscanbelongtomultiplerelations. Forexampleasegment item
can belong both tosegment relation and a sylstructure relation. Relations form an
ordered structure over the items within them. Items consist of a bundleof features
and a set of named linksto nodes in relation.
Figure3.1showsanexampleutterancewithasyntaxrelationandawordrelation.
The word relationcontains nodes that have next and previous connections whereas
the syntax relationhas up and down connections. Every node in the syntax tree is
CAT:S
CAT:VP
CAT:NP
name:”this”
CAT:pro
name:”is”
CAT:verb
name:”an”
CAT:art
name:”example”
CAT:noun
next
previous
Word Relation
Syntax Relation
Figure 3.1: An example representation of an utterance structure. This example
showsthewordrelationandthesyntaxrelation. Thesyntaxrelation(shownontop)
isa tree with linksconnecting the nodes, shown asblack circles. The word relation
(shown onthe bottom)isalist. The itemscontainthe actuallinguisticinformation
andareshownintheroundedboxes. Thedottedlinesshowtheconnectionsbetween
the nodes and items.
3.4 Relations
The relationsfor the basic English TTS are listed below.
Text Relation contains a single item which contains a feature with the input
character string that will besynthesized.
Token Relation is a list of trees where root of each tree contains tokenized
objectobtained fromthe inputcharacterstring. Punctuationandwhitespace
theTokenRelation,leavesofthePhrase Relationand rootsofthe Sylstructure
Relation.
Phrase Relation represents the words in the utterance. Words are leaves of
theTokenRelation,leavesofthePhrase Relationand rootsofthe Sylstructure
Relation.
Syllable Relationrepresentsa simplelistof syllable items which are the
inter-mediatenodes inthe Sylstructure Relation.
Segment Relation represents a simple list of segment(phoneme) items, which
formthe leaves of the Sylstructure Relationthrough which we can nd where
eachsegment is placed.
Sylstructure Relation represents alist of tree structures over the items in the
Word, Syllable and Segment items.
IntEvent Relation represents a simple list of intonation events (accents and
boundaries) whichare relatedtosyllables through the Intonation relation.
Intonation Relation represents a list of trees whose roots are items in the
Syllable Relation.
WaveRelationconsistsofasingleitemthathasafeaturewiththesynthesized
waveform.
Target Relation is a list of trees whose roots are segments and daughters are
F0target points.
3.5 Modules
The synthesis process in Festival consists of applying a number of modules to an
utterance. Each module will access various relations and items and generate new
features,itemsandrelations. Afterthemodulesareapplied,theutterancestructure
will be lled in and the waveform will be generated. The execution order of the
(Token_POS utt) (Token utt) (POS utt) (Phrasify utt) (Word utt) (Pauses utt) (Intonation utt) (PostLex utt) (Duration utt) (Int_Targets utt) (Wave_Synth utt) )
The modules used for TTS havethe following functions
Token Pos
Identies basic tokens, mainlyfor homographdisambiguation.
Token
Applies the token-to-word rules, and builds the Word Relation.
POS
Ifdesired, used as astandard part-of-speech tagger.
Phrasify
Buildsthe phrase relationusing the speciedmethod.
Word
Works as alexicallookup and builds the syllable and segment relations.
Pauses
Predicts pauses and inserts silence into the Segment Relation. (using
predic-tion mechanisms)
Intonation
Predictsaccentsand boundaries, buildsthe Intevent and Intonation relations
Organizes rules that can modify segments based on their context. (Used for
vowel reduction, contraction)
Duration
Predictsduration of segments.
Int Targets
Creates the Target Relation representing the desired F0contour.
Wave Synth
Calls the appropriate method togenerate the waveform.
3.6 Utterance Building
Utterance structures are usually used in the runtime process of converting text to
speech. However, one can use them also in database representation. The idea
behind this database representation is that we want to build utterance structures
for each utterance in a speech database. After obtainingthe utterance structures,
as if they had been correctly synthesized, one can use these structures for training
variousmodels. Forinstancegiventhe actualdurationsforthesegmentsinaspeech
database and utterance structures for these one can save the actual durations and
features (phonetic, prosodic context) which in uence the durations and training
models for that data.
In order to buildan utterance, we need labelles for the followingrelations.
Segment Labels: Segments must be labelledwith correct boundaries
consider-ingthe phone set of the language.
SyllableLabels: Syllablesmustbelabelledwithstressmarkingandboundaries
should bealigned with the segment boundaries.
Word Labels: Words must be labelled with boundaries aligned close to the
syllable and segment labels.
eachprosodicphrase.
TargetLabels: Target labelsare the meanF0values inHertzatthe mid-point
of eachsegment.
Among these labellings, segment labelling is the hardest to generate since any
au-tomatic method willhave tomake lowlevelphonetic classication. Computers are
not very good at this. Autoaligning is another alternative but afterwards hand
correction is necessary.
3.7 Diphone Databases
Adiphoneisasubwordunittypewhichcomprisetwophones. However,theduration
of a diphone is on the average one phoneme long since the beginning of a diphone
startsfromthemiddleoftherst phoneandtheend ofthediphone isatthe middle
of the second phone. The word hello can be mapped into the diphone sequence :
/sil-hh/,/hh-ax/, /ax-l/, /l-ow/, /ow-sil/.
For diphone synthesis, nearly allpossiblephone-phone transitionsina language
must be listed. In general, the number of diphones in a language is the square of
the number of phones. However due to phonotactic constraints some phone-phone
pairs may not occur at all. However people can often generate the non-existent
diphones if they try. Moreover, one must think about phone pairs that cross over
word boundaries as well. But even then certain combinations can not exist; for
example, /hh-ng/ diphone in English is probably impossible. The /ng/ phoneme
may only appear after the vowel in a syllable-initialposition. The /hh/ phoneme
can not appear at the end of a syllable, though sometimes it may be pronounced
when trying toadd aspirationtoopen vowels.
In fact,co-articulatoryeects may goovermore than twophones. However, the
diphone method assumes that this is not true. Unlike unit selection, which willbe
explainedlater, only one occurrence of a diphone is recorded. This makes selection
easier but collectiontask islaborious.
When humans are given a context that carries an unusual phoneme, they try
syn-Formant and articulatory synthesizers have advantage here. Since diphones need
to be cleanly articulated, various techniques have been proposed. One technique
is to use target words embedded carrier sentences to ensure that the diphones are
pronounced with acceptable duration and prosody (i.e consistently). One other
technique is using nonsense words that iterate through all possible diphone
com-binations. The advantage of using nonsense words is that the presentation is less
prone to pronounciation errors. For best results, the words should be pronounced
with consistentvocaleort, withas littleprosodicvariation aspossible.
Pronounc-ing in a monotone way is ideal. Nonsense words consist of a carrier where the
diphone is usually taken from a middle syllable. Classes of diphones, for instance:
vowel-consonant, consonant-vowel, vowel-vowel and consonant-consonant must be
extracted. Then, carrier contexts have tobe dened for these groups.
Thefollowingpseudocodelistsnonsensewordsforallpossiblelistofvowel-vowel
diphones.
for v1 in vowels
for v2 in vowels
print pause t aa t $v1 $v2 t aa pause
Onemustconsiderhoweasyitisforthespeakertopronouncethesenonsensewords.
3.7.1 Extracting the Pitchmarks
Festivalsupports residual excited Linear-Predictive-Coding (LPC) resynthesis [27].
It does support PSOLA [29] but this isnot distributed inthe public version. Both
of these techniques are pitchsynchronous, meaningthey require informationabout
where pitch periodsoccur inthe acoustic signal. If itis possible, recording with an
electroglottograph(EGG, alsoknown as laryngograph)isbetter. TheEGGrecords
electricalactivityinthe glottisduringspeech,whichmakesiteasiertogetthe pitch
moments.
Although extracting pitch periods from the EGG signal is not very easy, it is
fairly straightforward inpractice as the EdinburghSpeechToolsincludea program
it lls in the unvoiced section with the default pitchmarks. However it is not fully
automatic andrequires someone toinspect the resultand play withthe parameters
so toimprove the results.
If the signal is inverted -inv should be added to the arguments to pitchmark.
Theobjectistoproduceasinglemarkatthepeakofeachpitchperiodandphantom
periods duringunvoiced regions.
The command is as follows: Notice that -min and -max arguments are speaker
dependent.
pitchmark lar/file001.lar -o pm/file001.pm -otype est
-min 0.005 -max 0.012 -fill -def 0.01 -wave_end
If EGG signals do not exist for the diphones, another alternative is extracting the
pitchperiodsusingsomeothersignalprocessingfunction. Findingthepitchperiods
is similar to nding the F0 contour. It is harder than extracting from the EGG
signals but stillpossible with clean laboratory recorded speech.
The following script is a modication of the above script above for extracting
waveforms from a raw waveform signal. It is not as good as extracting the EGG
signal but it works. It is more computationally expensive since it requires rather
high order lters. The value should be changed according to the speaker's pitch
range.
for i in $
do
fname='basename $i. wav'
echo $i
$ESTDIR/bin/ch_wave -scaleN 0.9 $i -F 16000 -o /tmp/tmp$$.wav
$ESTDIR/bin/pitchmark /tmp/tmp$$.wav -o pm/$fname.pm
-otype est -min 0.005 -max 0.012 -fill -def 0.01
-wave_end -lx_lf 200 -lx_lo 71 -lx_hf 80 -lx_ho 71 -med_o 0
done
If the pitch periods are extracted automatically, it is worth taking more care to
emulabel tool. Figure 3.2 shows the pitchmarks extracted with the 'pitchmark'
command. The pitchmarks (vertical lines) should be aligned to the largest peak
(circles) in each pitch period. In this gure this aligning is satisfactory for the
pitchmarks.
Figure 3.2: Close-up pitchmarks inwaveform signal
3.8 Unit Selection Databases
Unitselectionisthe selectionofunitof speech whichmaybeanything fromawhole
phrase down to a diphone (or even smaller). Technically, diphone selection is a
simplecase of this. In unit selection, thereisusually more than oneexample of the
unitandsomemechanismisusedtoselectbetweenthematrun-time. Unitselection
startswithaphoneticandprosodicspecicationforadesiredutterance. Eachphone
hasafeaturevector,includingatleastpitch,duration,andstressandalsocarriesthe
phonetic context from its preceding and following phones. ATR's CHATR system
[28] is an excellent example for the method of selectingbetween multipleexamples
andspit theroundness ofthe followingvowel,aect thepronounciationof theletter
s although there is an intermediate stop. It is not only obvious segmental eects
that cause variation in pronounciation, syllable position, word/phrase initial and
nal position have dierent level of articulations. Inter-syllable or intra-syllable or
word-initialorword-internalpositionsaectarticulation. Stressing andaccentsalso
cause dierences. Rather than listing all of these events and recording all of them,
an alternative is to take a natural distribution of speech and (semi-)automatically
nd the distinctions that exist rather than predening them. The success of such
systems vary. They can produce very high quality, natural sounding synthesis.
However when the database has unexpected holes or the selection costs fail, they
can produce verybad synthesis too.
3.8.1 Cluster Unit Selection
This part is a reimplementation of the techniques described in Black and Taylor's
work[30]. Unitselectionisbasedontaking adatabase ofgeneralspeechand trying
to cluster each phone type into groups of acoustically similar units based on the
(non-acoustic)informationavailableatsynthesis time. Thesenon-acoustic
informa-tion include phonetic context, prosodicfeatures (F0 and duration), stressing, word
position and accents. This work is similar toprevious works like CHATR selection
algorithm[28]andthe workofDonovan[32]. But thiswork diersfromHunt'work
[28] since it builds CART (Classication and Regression Tree) trees to select the
appropriate clusterof candidatephones and as a result itdoes not calculatetarget
costs (through linearregression)atselectiontime. As theclusters are builtdirectly
fromthe acousticguresand targetfeatures, atarget estimationfunction isnot
re-quired. Thisclustering methoddiersfromDonovan's worksince ituses adierent
acousticcost function(Donovanuses HMM's). Donovanselectsonecandidatewhile
in this technique a group of candidates are selected. The basic processes involved
inbuilding awaveform synthesizer for clustering algorithmisas follows:
Collect the database of general speech.
chronous analysis (LPC)
Build distance tables
Dump selection features (phone context, prosodic, positional) for each unit
type.
Buildclustertreesusing'wagon'withthefeaturesandacousticdistancesdumped
by the previous two stages.
Build the voice description itself.
3.8.2 Diphones from general databases
In this method, we use the diphones as a unit. We should have a general database
that is labelled with utterances as described above. We can extract a standard
diphone database fromthis general database. However, we may be unableto cover
all phoneme-phoneme transitions. Even in phonetically rich databases likeTimit 2
,
somevowel-voweldiphones doesnot exist. Wecan extractadiphone database from
the general databases but there may be some holes.
3.9 Building prosodic models
3.9.1 Phrasing
Prosodicphrasingin speechsynthesis makes the speech moreunderstandable. Due
to our lungs, there is a nite length of of time we can talk without taking a new
breath. This denes the upper bound on prosodic phrases. However, we usually
take breath before this upper bound. We use phrasing to mark groups within the
speech.
For Englishand many other languages,simple rules that are based on
punctua-tionareverygoodpredictorsofprosodicphraseboundaries. Ifthereisapunctuation
than thereusually exists aprosodicboundary. But sometimesaprosodicboundary
exists althoughthere isnopunctuationmark. Thusaphrosodicphrasingalgorithm
adding punctuationatdesired phrase breaks is possible and adequate.
Festival supports two methods for predicting prosodic phrases. The rst basic
method is by CART (Classication and Regression Tree). A test is made on each
word to predict if it is at the end of a prosodicphrase. The CART (Classication
and Regression Tree) tree returns B(short break) or BB(long break). BB denotes
end of utterance.
The following tree adds a break after the last word of a token that has the
followingpunctuation. (set! simple_phrase_cart_tree ' ((lisp_token_end_punc in ("?" "." ":")) ((BB)) ((lisp_token_end_punc in ("'" "\"" ";" ",")) ((B))
((n.name is 0) ;; end of utterance
((BB))
((NB))))))
As the basic punctuation model underpredicts, we need information that will
ndreasonableboundarieswithinstringsofwords. InEnglish,boundaries aremore
likely between content words. If we have no data totrain from, then written rules
ina CART tree can givea phrasingmodelbetter than pure punctuationrules.
To implement such a scheme we need threebasic functions:
determining if the current word is a function or content word
determining number of words since previous punctuation
determining number of words to next punctuation
A much better method for predicting phrase breaks is using a full statistical
modeltrained fromdata. But the problemwith this methodis that youneed a lot
of trainingdata to train phrasebreak models.
Wagontool 3
can beused toproduce CART(ClassicationandRegression Tree)
3
lisp token end punc
lisp until punctuation
lisp since punctuation
p.gpos
gpos
n.gpos
However withouta good intonation and durationmodelspending time on
pro-ducinggood phrasingisprobably not worth it.
3.9.2 Accent/Boundary Assignment
Content words are the key words ofa sentence. They are the importantwords that
carrythemeaningorsense. Forexampleinthefollowingsentence,capitalizedwords
are content words.
Willyou SELLmy CAR because I'veGONE toFRANCE
The rest of the words are structure words. They make the sentence grammatically
correct. For English, the placements of accents on stressed syllables in all
con-tentwords is quiteareasonable approximation and achivesabout80% accuracy on
typical databases. Using this method achieving simple, in other words discourse
neutral intonationisrelativelyeasy. Butachievingrealistic,naturalaccent
place-mentisstillbeyond this method. Here isasimple CARTtree that predicts accents
of stressed syllablesin contentwords.
(set! simple_accent_cart_xtree
'
(
(R:SylStructure.parent.gpo s is content)
)
)
)
3.9.3 F0 Generation
Based on the place of the accents, an F0 contour must be built. Accent positions
in uence durations and the F0 contour can not be generated without knowing the
durationsofthe segmentsthecontour istobegeneratedover. Thereare threebasic
F0generationmodulesinFestival: bygeneralrule, bylinear regression/CART, and
by TILT[33] 4
.
3.9.4 F0 by rule
This is the most general F0 generation method. This method allows target points
to be programmaticallycreated for each syllable in the utterance. The simple idea
behind this generalmethodis thata LISPfunction iscalledfor eachsyllable inthe
utterance. ThisLISP functionreturnsa listoftarget F0pointsthat liewithinthat
syllable. This method allows the user to program any F0 value. The idea behind
this technique extends from the implementation of TOBI type accents, where a
number of points are predicted for each accent. The baseline is the average F0 of
the speaker. To the end of the phrase the F0 declines slowly. This technique and
TOBIplaceF0targetpointsaboveand belowthatbaselinedependingontheaccent
type and position in phrase. For example a simple accent can be generated using
this technique as follows.
(define (targ_func1 utt syl)
"(targ_func1 UTT STREAMITEM)
Returns a list of targets for the given syllable."
(let ((start (item.feat syl 'syllable_start))
(end (item.feat syl 'syllable_end)))
(if (equal? (item.feat syl "R:Intonation.daughter1. name ") "Accented")
(list
4
(list (/ (+ start end) 2.0) 140)
(list end 100 )))))
This method checks if the current syllable is accented and returns a list of target
pairs. It assigns 110 Hz at the start point, 140 Hz at the mid-point of the syllable
and nally100 Hz at the end of the syllable.
This technique can be expanded with otherrules as neccesary. Festival includes
animplementation of TOBI using this technique.
3.9.5 F0 by linear regression
This technique nds the appropriate F0 target value for each syllable based on
available features obtained from training data. A set of features are collected for
each syllable and a linear regression model is used to model three points on each
syllable. This technique provides reasonable synthesis and requires less analysis of
intonation models. In most synthesizers, the task of generating a prosodic tune
consists of two sub-tasks, the prediction of intonation labels (accents, tones, etc)
fromtextandthe generationofacontourfromthoselabels[17]. F0contourscanbe
generated from TOBI labelled utterances. The TOBI labelling system [17] oersa
method for labellingpertinent aspects of intonationin speech. Although, there are
recognized limitationswith the system, ithas been used tohand-labellarge speech
databasesandisbeingusedinanumberofsynthesissystems. TOBIlabellingforan
utterance consists of three tiers each related (through time) to a speech waveform
[17]. The tiers are: labels, break indices and miscellaneous. The label tier marks
pitchaccents,phrase accents and boundary tones. The break index tier marks one
of four levels of prosodic breaks. The miscellaneous tier may contain any other
labelling, such as background noise, coughing, laughing, dis uencies 5
or anything
elsethatmightbelabelled. OnemethodofgeneratinganF0contourfromsuchlabels
and breaks isdescribed in[20], whichis calledthe APL method. The APL method
predicts a number of target points for each syllable marked with a pitch accent,
phrase accent or boundary tone. A number of specic rules deal with each case
Time
F0
Figure3.3: TOBI Parameters
accent introducesthreetarget points,the rst atheightH1 abovethe reference line
at the start of the syllable, the second at height H2 at the start, and the third at
H2 atthe end of the syllable, there is similar target pointsfor L*. The parameters
H1, H2etc. are given as fractionsof TopVal 6
and BaseVal 7
above orbelow RefVal 8
, so there is some independence from absolute pitch range. Independently, RefVal,
TopVal and BaseVal may decrease over time to represent declination. However,
this technique depends solely on training data (TOBI labellings), some issues for
instance multiple accents on syllables, accent placement with respect to the vowel
are not captured. The previous technique allows specication of structure without
explicit training from data, on the other hand this technique imposes no structure
and dependsondata. Tiltmodelling,which willbedescribed next, triesto balance
these two extremes.
6
SizeinHertzaboverefvalformaximumsizedaccents(speaker-dependent).
7
SizeinHertzbelowrefvalforminimumsizedaccents(speaker-dependent).
8
A Tilt labelling for an utterance consists of an assignment of one of four basic
intonational events: pitch accents, boundary tones, connections and silence. An
intonationaleventisageneraltermforphonologicallysignicantintonationaleect.
Connections represent the parts of contours where there is nothingof intonational
signicance. Tilt modelling is still under development and not as mature as the
othermethods. AtiltparameterizationofanaturalF0contourcanbeautomatically
derivedfromawaveformandalabellingof accentplacements. Analabelisusedfor
accents,b forboundaries, cforconnections,andsil forsilence. Foreach alabelfour
continousparametersarefound: height,duration,peakposition,withrespect tothe
vowel start, and tilt. This method gives better results when compared with linear
regression models but has not been tried on new languages other than English.
The automatic parameterization of apitch event on asyllable is interms of:
starting F0value(Hz)
duration
amplitude of rise (Arise, in Hz)
amplitude of rise (Arise, in Hz)
starting point, time aligned with the signal and with the vowel onset
Figure 3.4 shows the Tiltparamters.
Tiltparameter is the dierenceof the amplitudedivided by their sum.
tilt =
jArisej jAfallj
jArisej+jAfallj
The tiltparameter has arange of-1 to1where -1ispure fall,1 ispure rise and
0 contains equal portions of rise and fall.
3.9.7 Duration
Similar tothe prosody generation, simplesolutions for predicting durations of