抄録
Wesetarticulatoryclassesbasedonarticulatoryfeature,andweuseposteriorprobabilitiesonarticula-toryclassesforlanguageidenti cation.PosteriorprobabilityoneacharticulatoryclassiscalculatedbyarticulatoryfeatureextractorbasedonGMMs.Theposteriorprobabilityvaluesofthearticulatoryclassesareconcatenatedtoformvectorateachanalysisframe.ThesevectorsarethenquantizedtoyieldVQcodesequence,whichisusedasthetrainingdataforan-gramlanguagemodel.Theresultsoflanguageidenti cationexperimentbetweenJapaneseandEnglishshowedchangeinidenti cationperformancebycodebooksize.Themethodthatuseslanguage-dependentarticulatoryfeatureextractorshowedidenti cationrateof98.1%whencodebooksizewas64,andthemethodthatuseslanguage-independentarticulatoryfeatureextractorshowedidenti cationrateof95.6%whencodebooksizewas256.