自动语音识别中声学模型鉴别性训练的研究与应用.pdfVIP

自动语音识别中声学模型鉴别性训练的研究与应用

Abstract

Withthedevelopmentofartificialintelligencetechnology,

automaticspeechrecognition(ASR)hasbecomeapopularapplication

inrecentyears.However,therearestillchallengesforASRsystems,

suchasrecognizingspeechundernoisyenvironments,differentaccents

anddialects.InordertoimprovetheperformanceofASRsystems,the

discriminativetrainingofacousticmodelshasattractedincreasing

attention.Thispaperdiscussestheresearchandapplicationsof

discriminativetrainingforacousticmodelsinASR.

Introduction

Automaticspeechrecognition(ASR)iswidelyusedinmanyfields

withtherapiddevelopmentofartificialintelligencetechnology.ASR

technologyhasmadesignificantprogressinrecentyears,however,the

recognitionrateofcurrentmainstreamsystemsstillhasroomfor

improvement,especiallyundernoisyenvironments,differentaccents

anddialects.

IntraditionalASRsystems,theacousticmodelsaretrainedusing

maximumlikelihoodestimation(MLE)andareoptimizedbasedonthe

likelihoodcriterion.However,theMLE-basedapproachassumesthat

theacousticmodelsareequallywellonallspeechsounds,whichcan

limittheirgeneralizationperformanceinreal-worldsituations.To

addressthisissue,discriminativetrainingofacousticmodelshasbeen

proposed.

DiscriminativeTrainingofAcousticModels

Discriminativetrainingofacousticmodelsisbasedontheprinciple

ofmaximummutualinformation,whichaimstominimizethedifference

betweentheoutputofacousticmodelsandthereferencetranscription.

Thegeneralideaistooptimizetheacousticmodelstomakethemmore

discriminativeandbetterabletodifferentiatebetweenspeechsounds.

Onepopulardiscriminativetrainingapproachismaximummutual

information(MMI)criterion,whichconsidersnotonlythelikelihoodof

speech,butalsothelikelihoodofotherspeechhypothe