资源描述:
《MODELLING PRONUNCIATION VARIATION SOME PRELIMINARY RESULTS》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、MODELLINGPRONUNCIATIONVARIATION:SOMEPRELIMINARYRESULTSMirjamWester,JudithKessens,CatiaCucchiarini,HelmerStrikIn:H.Strik,N.Oostdijk,C.CucchiariniandP.A.Coppen(eds.)ProceedingsoftheDepartmentofLanguage&SpeechVol.20,pp.127-137,Nijmegen,theNetherlands,1997MODELLINGPRONUN
2、CIATIONVARIATION:SOMEPRELIMINARYRESULTSMirjamWester,JudithKessens,CatiaCucchiarini,HelmerStrikAbstractInthispaperwedescribeamethodforimprovingtheperformanceofacontinuousspeechrecognizerbymodellingpronunciationvariation.Althoughtheresultsobtainedwiththismethodareinlin
3、ewiththosereportedbyotherauthors,themagnitudeoftheimprovementsisverysmall.Inlookingforpossibleexplanationsfortheseresults,wecomputedvarioussortsofstatisticsaboutthematerial.Sincethesedataprovedtobeveryusefulinunderstandingtheeffectsofourmethod,theyarediscussedinthisp
4、aper.Moreover,onthebasisofthesestatisticswediscusshowthesystemcanbeimprovedinthefuture.1.IntroductionAttheDepartmentofLanguageandSpeechoftheUniversityofNijmegenweareworkingonaSpokenDialogueSystem(SDS)thatwillbeemployedtoautomatepartofapublictransportinformationservic
5、e.ThissystemwasadaptedfromaGermanprototypedevelopedbyPhilipsResearchLabs,andwasfurtherimprovedbymeansofabootstrappingmethod(Striket.al.,1996and1997).AnimportantcomponentofthisSDSisacontinuousspeechrecognizer(CSR).ThispartoftheSDSwasalsograduallyimprovedthroughtheboot
6、strappingmethod,byaddingmoredata.However,sinceapointwasreachedatwhichnofurtherincreaseinperformancecouldbeobtainedbyincreasingthedata,newmethodsofimprovingthesystemweresought.GiventhattheSDSisamixed-initiativesystemandthatthekindofspeechthecallersmayuseisextremelyvar
7、ied,wethoughtofimprovingthesystem’sperformancebymodellingpronunciationvariation.Inthispaper,themethodusedformodellingpronunciationvariationisdiscussedindetailinsection2.Subsequently,insection3theresultsobtainedwiththismethodarepresentedtogetherwithvarioussortsofstati
8、sticsaboutthematerial.Insection4wediscusshowthestatisticswecomputedhelpedustounderstandwhythevariationsinperformanceweresosmall,and