搜索
5725--IAbstractUnderthedevelopmentofinformationglobalization,peopleneedtext-to-speechtechniquemoreandmore.Thishasmadeoutstandingimprovementinresearchanddevelopmentoftext-to-speechtechnique.Althoughthequalityofsynthesizedspeechhasbeenaccepted,thenaturalnessisstilldissatisfying.Thispaperstartswithpartofthefrontendtechnique,whicharetext-to-syllable,detectionofprosodyphraseboundariesandspeechstressdetermining,inordertosolvethepronunciationerrorsandtheimproperhandlinginrhythm,pause,stressduringtexttospeech.Sothesynthesizedspeechoftext-to-speechsystemcanbemademorenatural.Afteranalyzingtheshortcomingsofdealingwithpolyphonesinmethodsbasedonruleandonstatistics,weintroduceacomparativelyself-containedflow,whicharethreesteps:lookinthepolyphonelist,fitthepartofspeechrulesandcomputesemanticsimilarities.ThispaperdescribesHow-Netindetailandregardsitasthesettingofsemanticcomputationanddeductioninordertocomputethewordsimilarityandthecontextwindowsimilarity.Theexperimentresultsshowthatthiskindofmethodfordealingwithpolyphoneshasmanyvirtues:usingsemanticcomputingcanfindthesemanticsimilarwordoftheunknownpronunciationword,andthengetthepronunciationofpolyphonesinitfromitssimilarword.Sothismethodcanimprovenotonlytheaccuracyofdeterminingpolyphonepronunciationsgreatlybutalsotheaccuracyoftext-to-syllableintext-to-speechsystem.Thoughsomeresearchershavegotsomegoodexperimentresultsindetectionofprosodyphraseboundaries,theirmethodsareallverycomplex.Inthispaperwebuildaself-learningsystemfordetectionrulesofprosodyphraseboundariesusingthetransformation-basederror-drivenlearning,whichhassucceededinparsing.Afterbuildingalargecorpusthathas5725textandspeechsentences,theself-learningsystemretrievesalargeamountofrules.Notonlyretrievingoftheserulesiseasytousebutalsotheaccuracyofautomaticdetectionofprosodyphraseboundariescanberaisedwhenusingtheserules.In--IIAbstractordertoreducetheburdenbroughtbythelargequantityofrulesinTTSsystem,thispaperusestreetoorganizerules.Theautomaticdetectionofprosodyphraseboundarieshasbeenusedwellinrealsystem.WhetherthewordstressformulationexistsinChinesemandarinhasnotgotthelastword,butresearchoftheChinesemandarinstressisstillveryimportanttotext-to-speechsystem.Sothispaperstartswiththeprincipleofproducingandapperceiveofprosodywordstressandanalyzestherulesoftwosyllablewords’stress.Thenusingthesameself-learningsystemasdetectionofprosodyphraseboundariestogetrulesofprosodyphrases’stress.Buttheexperimentresultsarenotgoodenoughtobeusedintherealsystem.Wemustresearchitfurtherinthefuture.Keywordspolyphone;semanticsimilarity;prosodyphrase;prosodyword;stress--III......................................................................................................................IAbstract..............................................................................................................II1.........................................................................................................11.1........................................................................................11.2....................................................................................21.3.................................................................31.4............................................................................................51.5.........................................................................62.....................................................................................72.1................................................................................................72.2....................................................................................................82.2.1.................................................................................82.2.2.................................................................................92.2.3................................................................112.2.4...............................................................................112.3.......................................................................142.3.1...........................................................................142.3.2.......................................................................................142.3.3...............................................................................162.3.4................................................................172.4......................................................................................182.5..................................................................................................203...............................................................................213.1..................................................................................................213.2..................................................................................213.2.1........................................................233.2.2...................................................................................243.