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.