Proceedings Fonetik 2009 - Institutionen för lingvistik

More documents

Recommendations

Info

Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm UniversityAdapting the Filibuster text-to-speech system forNorwegian bokmålKåre Sjölander and Christina TånnanderThe Swedish Library of Talking Books and Braille (TPB)AbstractThe Filibuster text-to-speech system is specificallydesigned and developed for the productionof digital talking textbooks at universitylevel for students with print impairments. Currently,the system has one Swedish voice,'Folke', which has been used in production atthe Swedish Library of Talking Books andBraille (TPB) since 2007. In August 2008 thedevelopment of a Norwegian voice (bokmål)started, financed by the Norwegian Library ofTalking Books and Braille (NLB). This paperdescribes the requirements of a text-to-speechsystem used for the production of talking textbooks,as well as the developing process of theNorwegian voice, 'Brage'.IntroductionThe Swedish Library of Talking Books andBraille (TPB) is a governmental body that providespeople with print impairments with Brailleand talking books. Since 2007, the in-housetext-to-speech (TTS) system Filibuster with itsSwedish voice 'Folke' has been used in the productionof digital talking books at TPB(Sjölander et al., 2008). About 50% of theSwedish university level textbooks is currentlyproduced with synthetic speech, which is afaster and cheaper production method comparedto the production of books with recorded humanspeech. An additional advantage is that thestudent gets access to the electronic text, whichis synchronized with the audio. All tools andnearly all components in the TTS system componentsare developed at TPB.In August 2008, the development of a Norwegianvoice (bokmål) started, financed by theNorwegian Library of Talking Books andBraille (NLB). The Norwegian voice 'Brage'will primarily be used for the production of universitylevel textbooks, but also for news textand the universities' own production of shorterstudy materials. The books will be produced asDAISY-books, the international standard fordigital talking books, via the open-sourceDAISY Pipeline production system (DAISYPipeline, 2009).The Filibuster system is a unit selectionTTS, where the utterances are automaticallygenerated through selection and concatenationof segments from a large corpus of recordedsentences (Black and Taylor, 1997).An important feature of the Filibuster TTSsystem is that the production team has totalcontrol of the system components. An unlimitednumber of new pronunciations can be added, aswell as modifications and extensions of the textprocessing system and rebuilding of the speechdatabase. To achieve this, the system must beopen and transparent and free from black boxes.The language specific components such asthe pronunciation dictionaries, the speech databaseand the text processing system are NLB’sproperty, while the language independent componentsare licensed as open-source.Requirements for a narrative textbooktext-to-speech systemThe development of a TTS system for the productionof university level textbooks calls forconsiderations that are not always required for aconventional TTS system.The text corpus should preferably consist oftext from the same area as the intended productionpurpose. Consequently, the corpus shouldcontain a lot of non-fiction literature to covervarious topics such as religion, medicine, biology,and law. From this corpus, high frequencyterms and names are collected and added to thepronunciation dictionary.The text corpus doubles as a base for constructionof recording manuscripts, which in additionto general text should contain representativenon-fiction text passages such as bibliographicand biblical references, formulas andURL’s. A larger recording manuscript thanwhat is conventionally used is required in orderto cover phone sequences in foreign names,terms, passages in English and so on. In addition,the above-mentioned type of textbook specificpassages necessitates complex and welldevelopedtext processing.36
Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm UniversityThe number of out-of-vocabulary (OOV)words is likely to be high, as new terms andnames frequently appear in the textbooks, requiringsophisticated tools for automatic generationof pronunciations. The Filibuster systemdistinguishes between four word types; propernames, compounds and simplex words in thetarget language, and English words.In order to reach the goal of making thetextbooks available for studies, all text - plainNorwegian text and English text passages, OOVwords and proper names - need to be intelligible,raising the demands for a distinct and pragmaticvoice.The development of the NorwegianvoiceThe development of the Norwegian voice canbe divided into four stages: (1) adjustments andcompletion of the pronunciation dictionary andthe text corpus, and the development of the recordingmanuscripts, (2) recordings of theNorwegian speaker, (3) segmentation and buildingthe speech database, and (4) quality assurance.Pronunciation dictionariesThe Norwegian HLT Resource Collection hasbeen made available for research and commercialuse by the Language Council for Norwegian(http://www.sprakbanken.uib.no/). The resourcesinclude a pronunciation dictionary forNorwegian bokmål with about 780,000 entries,which were used in the Filibuster NorwegianTTS. The pronunciations are transcribed in asomewhat revised SAMPA, and follow mainlythe transcription conventions in Øverland(2000). Some changes to the pronunciationswere done, mainly consistent adaptations to theNorwegian speaker's pronunciation and removalof inconsistencies, but a number of true errorswere also corrected, and a few changes weremade due to revisions of the transcription conventions.To cover the need for English pronunciations,the English dictionary used by the Swedishvoice, consisting of about 16,000 entries,was used. The pronunciations in this dictionaryare ‘Swedish-style’ English. Accordingly, theywere adapted into ‘Norwegian-style’ Englishpronunciations. 24 xenophones were implementedin the phoneme set, of which about 15have a sufficiently number of representations inthe speech database, and will be used by theTTS system. The remaining xenophones will bemapped into phonemes that are more frequentin the speech database.In addition, some proper names from theSwedish pronunciation dictionary were adaptedto Norwegian pronunciations, resulting in aproper name dictionary of about 50,000 entries.Text corpusThe text corpus used for manuscript constructionand word frequency statistics consists ofabout 10.8 million words from news and magazinetext, university level textbooks of differenttopics, and Official Norwegian Reports(http://www.regjeringen.no/nb/dok/NOUer.html?id=1767). The text corpus has been cleanedand sentence chunked.Recording manuscriptsThe construction of the Norwegian recordingmanuscript was achieved by searching phoneticallyrich utterances iteratively. While diphoneswas used as the main search unit, searches alsoincluded high-frequency triphones and syllables.As mentioned above, university level textbooksinclude a vast range of different domainsand text types, and demands larger recordingmanuscripts than most TTS systems in order tocover the search units for different text typesand languages. Biographical references, for example,can have a very complex construction,with authors of different nationalities, name initialsof different formats, titles in other languages,page intervals and so on. To maintain ahigh performance of the TTS system for morecomplex text structures, the recording manuscriptmust contain a lot of these kinds of utterances.To cover the need of English phone sequences,a separate English manuscript was recorded.The CMU ARCTIC database forspeech synthesis with nearly 1,150 English utterances(Kominek and Black, 2003) was usedfor this purpose. In addition, the Norwegianmanuscript contained many utterances withmixed Norwegian and English, as well as emailaddresses, acronyms, spelling, numerals, lists,announcements of DAISY specific structuressuch as page numbers, tables, parallel text andso on.RecordingsThe speech was recorded in NLB’s recordingstudio. An experienced male textbook speakerwas recorded by a native supervisor. The re-37
Page 1 and 2: Department of LinguisticsProceeding
Page 3 and 4: Proceedings, FONETIK 2009, Dept. of
Page 35: Proceedings, FONETIK 2009, Dept. of
Page 41 and 42: Proceedings, FOETIK 2009, Dept. of
Page 87 and 88:
Proceedings, FONETIK 2009, Dept. of
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

Create successful ePaper yourself

Delete template?

Save as template?