inAcousticModelRefinement JáchymKolá ř,Lud ěkMüller

Abstract: Automatic speech recognition (ASR) systems usually consist of an acoustic model and a languagemodel. This paper describes a techniqueof an efficient deployment of t heacousticmodel parameters. The acoustic model typically utilizes Continuous Density Hi dden Markov Models (CDHMM). The output probability of a particular CDHMM state is represe nted by a Gaussian mixturedensitywithadiagonalcovariancestructure.Usually,theoutputprobabil itydensityfunction ofeachCDHMMstatecontainsthesamenumberofmixturecomponentsalthougha differentnumber of components in individual statesmay yieldmore accurate recognition resul ts, especially for low-resourceASRsystems.Thecentralideaistoassignmorecomponentsto stateswhereitiseffectiveand lesscomponentstostateswheretheincreasingnumberofcomponentsisnot warrantingasignificantly betterdescriptionofthetrainingdata.Thenumberofmixturecomponentsfor aparticularCDHMM stateischosenbyoptimizingtheBayesianInformationCriterion(BIC)