Proceedings Fonetik 2009 - Institutionen för lingvistik

More documents

Recommendations

Info

Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm UniversityExperiments with Synthesis of Swedish DialectsBeskow, J. and Gustafson, J.Department of Speech, Music & Hearing, School of Computer Science & Communication, KTHAbstractWe describe ongoing work on synthesizingSwedish dialects with an HMM synthesizer. Aprototype synthesizer has been trained on alarge database for standard Swedish read by aprofessional male voice talent. We have selecteda few untrained speakers from each ofthe following dialectal region: Norrland, Dala,Göta, Gotland and South of Sweden. The planis to train a multi-dialect average voice, andthen use 20-30 minutes of dialectal speech fromone speaker to adapt either the standard Swedishvoice or the average voice to the dialect ofthat speaker.IntroductionIn the last decade, most speech synthesizershave been based on prerecorded pieces ofspeech resulting in improved quality, but withlack of control in modifying prosodic patterns(Taylor, 2009). The research focus has been directedtowards how to optimally search andcombine speech units of different lengths.In recent years HMM based synthesis hasgained interest (Tokuda et al., 2000). In this solutionthe generation of the speech is based on aparametric representation, while the graphemeto-phonemeconversion still relies on a largepronunciation dictionary. HMM synthesis hasbeen successfully applied to a large number oflanguages, including Swedish (Lundgren,2005).Dialect SynthesisIn the SIMULEKT project (Bruce et al., 2007)one goal is to use speech synthesis to gain insightinto prosodic variation in major regionalvarieties of Swedish. The aim of the presentstudy is to attempt to model these Swedish varietiesusing HMM synthesis.HMM synthesis is an entirely data-drivenapproach to speech synthesis and as such itgains all its knowledge about segmental, intonationaland durational variation in speech fromtraining on an annotated speech corpus. Giventhat the appropriate features are annotated andmade available to the training process, it ispossible to synthesize speech with high quality,at both segmental and prosodic levels. Anotherimportant feature of HMM synthesis, thatmakes it an interesting choice in studying dialectalvariation, is that it is possible to adapt avoice trained on a large data set (2-10 hours ofspeech) to a new speaker with only 15-30 minutesof transcribed speech (Watts et al., 2008).In this study we will use 20-30 minutes of dialectalspeech for experiments on speaker adaptionof the initially trained HMM synthesisvoice.Data descriptionThe data we use in this study are from the NorwegianSpråkbanken. The large speech synthesisdatabase from a professional speaker ofstandard Swedish was recorded as part of theNST (Nordisk Språkteknologi) synthesis development.It was recorded in stereo, with thevoice signal in one channel, and signal from alaryngograph in the second channel.The corpus contains about 5000 read sentences,which add up to about 11 hours ofspeech. The recordings manuscript was basedon NST’s corpus, and the selection was done tomake them phonetically balanced and to ensurediphone coverage. The manuscripts are not prosodicallybalanced, but there are different typesof sentences that ensure prosodic variation, e.g.statements, wh-questions, yes/no questions andenumerations.The 11 hour speech database has beenaligned on the phonetic and word levels usingour Nalign software (Sjölander & Heldner,2004) with the NST dictionary as pronunciationdictionary. This has more than 900.000 itemsthat are phonetically transcribed with syllableboundaries marked. The text has been part-ofspeechtagged using a TNT tagger trained onthe SUC corpus (Megyesi, 2002).From the NST database for training ofspeech recognition we selected a small numberof unprofessional speakers from the followingdialectal areas: Norrland, Dala, Göta, Gotlandand South of Sweden. The data samples areconsiderably smaller than the speech synthesisdatabase: they range from 22 to 60 minutes,28
Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm Universitycompared to the 11 hours from the professionalspeaker.HMM Contextual FeaturesThe typical HMM synthesis model (Tokuda etal., 2000) can be decomposed into a number ofdistinct layers:• At the acoustic level, a parametricsource-filter model (MLSA-vocoder) isresponsible for signal generation.• Context dependent HMMs, containingprobability distributions for the parametersand their 1 st and 2 nd order derivatives,are used for generation of controlparameter trajectories.• In order to select context dependentHMMs, a decision tree is used, that usesinput from a large feature set to clusterthe HMM models.In this work, we are using the standard modelfor acoustic and HMM level processing, andfocus on adapting the feature set for the decisiontree for the task of modeling dialectal variation.The feature set typically used in HMM synthesisincludes features on segment, syllable,word, phrase and utterance level. Segment levelfeatures include immediate context and positionin syllable; syllable features include stress andposition in word and phrase; word features includepart-of-speech tag (content or functionword), number of syllables, position in phraseetc., phrase features include phrase length interms of syllables and words; utterance levelincludes length in syllables, words and phrases.For our present experiments, we have alsoadded a speaker level to the feature set, sincewe train a voice on multiple speakers. The onlyfeature in this category at present is dialectgroup, which is one of Norrland, Dala, Svea,Göta, Gotland and South of Sweden.In addition to this, we have chosen to add tothe word level a morphological feature statingwhether or not the word is a compound, sincecompound stress pattern often is a significantdialectal feature in Swedish (Bruce et al.,2007). At the syllable level we have added explicitinformation about lexical accent type (accentI, accent II or compound accent).Training of HMM voices with these featuresets is currently in progress and results will bepresented at the conference.AcknowledgementsThe work within the SIMULEKT project isfunded by the Swedish Research Council 2007-2009. The data used in this study comes fromNorsk Språkbank (http://sprakbanken.uib.no)ReferencesBruce, G., Schötz, S., & Granström, B. (2007).SIMULEKT – modelling Swedish regionalintonation. Proceedings of Fonetik, TMH-QPSR, 50(1), 121-124.Lundgren, A. (2005). HMM-baserad talsyntes.Master's thesis, KTH, TMH, CTT.Megyesi, B. (2002). Data-Driven SyntacticAnalysis - Methods and Applications forSwedish. Doctoral dissertation, KTH, Departmentof Speech, Music and Hearing,KTH, Stockholm.Sjölander, K., & Heldner, M. (2004). Wordlevel precision of the NALIGN automaticsegmentation algorithm. In Proc of TheXVIIth Swedish Phonetics Conference, Fonetik2004 (pp. 116-119). Stockholm University.Taylor, P. (2009). Text-To-Speech Synthesis.Cambridge University Press.Tokuda, K., Yoshimura, T., Masuko, T., Kobayashi,T., & Kitamura, T. (2000). Speechparameter generation algorithms for hmmbasedspeech synthesis. In Proceedings ofCASSP 2000 (pp. 1315-1318).Watts, O., Yamagishi, J., Berkling, K., & King,S. (2008). HMM-Based Synthesis of ChildSpeech. Proceedings of The 1st Workshopon Child, Computer and Interaction.29
Page 1 and 2: Department of LinguisticsProceeding
Page 3 and 4: Proceedings, FONETIK 2009, Dept. of
Page 27: Proceedings, FONETIK 2009, Dept. of
Page 41 and 42: Proceedings, FOETIK 2009, Dept. of
Page 79 and 80:
Proceedings, FONETIK 2009, Dept. of
Page 81 and 82:
Page 83 and 84:
Page 85 and 86:
Page 87 and 88:
Page 89 and 90:
Page 91 and 92:
Page 93 and 94:
Page 95 and 96:
Page 97 and 98:
Page 99 and 100:
Page 101 and 102:
Page 103 and 104:
Page 105 and 106:
Page 107 and 108:
Page 109 and 110:
Page 111 and 112:
Page 113 and 114:
Page 115 and 116:
Page 117 and 118:
Page 119 and 120:
Page 121 and 122:
Page 123 and 124:
Page 125 and 126:
Page 127 and 128:
Proceedings, FOETIK 2009, Dept. of
Page 129 and 130:
Page 131 and 132:
Page 133 and 134:
Page 135 and 136:
Page 137 and 138:
Page 139 and 140:
Page 141 and 142:
Page 143 and 144:
Page 145 and 146:
Page 147 and 148:
Page 149 and 150:
Page 151 and 152:
Page 153 and 154:
Page 155 and 156:
Page 157 and 158:
Page 159 and 160:
Page 161 and 162:
Page 163 and 164:
Page 165 and 166:
Page 167 and 168:
Page 169 and 170:
Page 171 and 172:
Page 173 and 174:
Page 175 and 176:
Page 177 and 178:
Page 179 and 180:
Page 181 and 182:
Page 183 and 184:
Page 185 and 186:
Page 187 and 188:
Page 189 and 190:
Page 191 and 192:
Page 193 and 194:
Page 195 and 196:
Page 197 and 198:
Page 199 and 200:
Page 201 and 202:
Page 203 and 204:
Page 205 and 206:
Page 207 and 208:
Page 209 and 210:
Page 211 and 212:
Page 213 and 214:
Page 215 and 216:
Page 217 and 218:
Page 219 and 220:
Page 221 and 222:
Page 223 and 224:
Page 225 and 226:
Page 227:
Department of LinguisticsPhonetics
show all

Proceedings Fonetik 2009 - Institutionen för lingvistik

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?