上传者: longwo888
|
上传时间:2025/6/5 8:21:06
|
文件大小:617KB
|
文件类型:pdf
TheApplicationofHiddenMarkovModelsinSpeechRecognition
HiddenMarkovModels(HMMs)provideasimpleandeffectiveframeworkformodellingtime-varyingspectralvectorsequences.Asaconsequence,almostallpresentdaylargevocabularycontinuousspeechrecognition(LVCSR)systemsarebasedonHMMs.WhereasthebasicprinciplesunderlyingHMM-basedLVCSRareratherstraightforward,theapproximationsandsimplifyingassumptionsinvolvedinadirectimplementationoftheseprincipleswouldresultinasystemwhichhaspooraccuracyandunacceptablesensitivitytochangesinoperatingenvironment.Thus,thepracticalapplicationofHMMsinmodernsystemsinvolvesconsiderablesophistication.TheaimofthisreviewisfirsttopresentthecorearchitectureofaHMM-basedLVCSRsystemandthendescribethevariousrefinementswhichareneededtoachievestate-of-the-artperformance.Theserefinementsincludefeatureprojection,improvedcovariancemodelling,discriminativeparameterestimation,adaptationandnormalisation,noisecompensationandmulti-passsystemcombination.ThereviewconcludeswithacasestudyofLVCSRforBroadcastNewsandConversationtranscriptioninordertoillustratethetechniquesdescribed.
本软件ID:11029043