automatic learning of finite state automata for pronunciation modeling

《automatic learning of finite state automata for pronunciation modeling》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、AUTOMATICLEARNINGOFFINITESTATEAUTOMATAFORPRONUNCIATIONMODELING.M.PastorandF.CasacubertaInstitutTecnol`ogicd'Inform`aticaUniversitatPolit`ecnicadeVal`enciaCam´ıdeVeras/n46071Val`enciafmoises,fcng@iti.upv.esAbstractoftheacousticscoreoftheword.Shortwords(pronouns,articles,etc.)arethemostproblemati

2、cones,asmallvari-Thegreatvariabilityofwordpronunciationsinsponta-ationwithrespecttothecanonicalpronunciationcanrep-neousspeechisoneofthereasonsforthelowperfor-resentanimportantvariationwithrespecttothecanonicalmanceofthepresentspeechrecognitionsystems.Therepresentation.Thesewordsarecommoninhuma

3、nlan-generationofdictionariesthattakeintoaccountthisvari-guageandusuallydonothavesemanticinformation,andabilitycanincreasetherobustnessofsuchsystems.Ahaveahighn-gramprobabilityofocurrence.Thisisway,wordpronunciationisapossiblephonesequencethatforhumans,anaccuratepronunciationofthesewordsiscanap

4、pearinarealutterance,andrepresentsapossiblenotimportantinordertobeunderstood.[11].acousticrealizationoftheword.Thereareseveralapproachestoautomaticpronuncia-Inthispaper,wordpronunciationsaremodeledusingtionmodeling.Oneofthemostinterestingapproachesisfinitestateautomata.Theuseofsuchmodelsallowfor

5、thephoneme-basedrule-learningtechnique[10,11].Thetheapplicationofgrammaticalinferencemethodsandanmainproblemthatariseswiththesetechniquesisover-easyintegrationwiththeotherssourcesofacknowledge.generalization.Ontheotherhand,oneoftheirgreatestThetrainingsamplesareobtainedfromthealignmentbe-advant

6、agesistheireasyextensiontoinfrequentorun-tweenthephonedecodificationofeachtrainingutteranceobservedwords.However,ifawordisinfrequent,theandthecorrespondingcanonicaltranscription.effectontheglobalperformanceofthesystemissmall.ModelsproposedinthisworkwereappliedinaForthesewords,weuseonlythecanonic

7、alpronuncia-translation-orientedspeechtask.Theimprovementstionasamodelasisusedforlongwords.Anotheral-achievedbythesenewmodelswereintherangebetwenternativeistousefinitestateautomataaspronunciation2.7to0.6pointsdependingonthelanguage