Thepronunciationvariabilityisanimportantissuethatmustbefacedwithwhendevelopingpracticalautomaticspontaneousspeechrecognitionsystems.Inthispaper,thefactorsthatmayaffecttherecognitionperformanceareanalyzed,inculdingthosespecifictotheChineselanguage.BystudyintheINITIAL/FINAL（IF）characteristicsofChineselanguageanddevelopingtheBayesianequation,theconceptsofgeneralizedINITIAL/FINAL（GIF）andgeneralizedsyllable(GS),theGIFmodelingandtheIF-GIFmodeling,aswellasthecontext-dependentpronunciationweighting,areproposedbasedonawellphoneticallytranscribedseeddatabase.Byusingthesemethods,theChinesesylableerrorrate(SER)isreducedby6.3%and4.2%comparedwiththeGIFmodelingandIFmodelingrespectivelywhenthelanguagemodel,suchassyllableorwordN-gram,isnotused.Theeffectivenessofthesemethodsisalsoprovedwhenmoredatawithoutthephonetictranscriptionareusedtorefinetheacousticmodelusingtheproposediterativeforced-alignmentbasedtranscribing(IFABT)method,achievinga5.7%SERreduction.
