Applying PCFG in Parsing the N-best Speech Lattice directly in 在分析信息的N-最好的语音格直接应用

Using PCFG in __ndarin Continuous Speech RecognitionGuoDong Zhou and Kim Teng LuaDepartment of Com

1 UsingPCFGin__ndarinContinuousSpeechRecognition GuoDongZhouandKimTengLua DepartmentofComputerScien__ SchoolofComputing NationalUniversityofSingapore LowerKentRidgeRoad Singapore119260 {zhougd,luakt}@comp.nus.edu.sg Abstract: Thispaperdescribesthecon__ptofprobabilisticcontext-free gram__r(PCFG)andpresentsanefficientchartparsingalgorithmwhichcan directlydecodethe__ndarinspeechlatti__.Intheso-calledN-bestBST(base syllable+tone)speechlatti__,notonlyaretheN-bestpaths(basesyllable sequen__swithdifferentsegmentation)kept,buttheN-bestbasesyllable candidatesarealsoreservedforeverybasesyllablesegmentinthespeechlatti__.It isfoundthatdecodingtheN-bestBSTspeechlatti__achievesbetterperfor__n__ thandecodingthetop-bestspeechlatti__.Inordertobetteresti__tetherule probabilitiesinPCFG,amixedtrainingmethodisproposed.Itisfoundthatthe PCFGusingthemixedtrainingmethodhasbetterperfor__n__thanusing supervisedorunsupervisedtraining.ItisconcludedthatthePCFGcanbeusedasa powerfullanguagemodelforspeechrecognitionandthatitispreferabletodecode theN-bestBSTspeechlatti__insteadofthetop-bestBSTspeechlatti__,largely duetoinsertionanddeletionproblems. Keywords: __ndarinspeechrecognition,Probabilisticcontext-freegram__r, Chartparsingalgorithm,B-bestBSTspeechlatti__,MixedtrainingofPCFG. 1.Introduction Naturallanguagecanbeviewedasastochasticpro__ss.Inspeechrecognition,an acousticsignalisgivenandthegoalistofindthelinguistichypothesisthat __ximizes.UsingBayesLaw: (1) speech AsshowninFigure1,foragivensignal,isesti__tedbya recognizer ,whichcomparestoitsstoredacousticmodelsofallspeechunits. Providinganesti__teforistheresponsibilityofacousticmodelingwhile providinganesti__teforistheresponsibilityoflanguagemodelingthrougha languagedecoder .