Mandarin phonetic recognition using mixture hidden Markov models with time duration function

Abstract

This paper proposes mixture hidden Markov models (HMM) with a time duration function to solve the recognition of Mandarin Chinese diphthongs and several words that resemble diphthongs. We propose an autoregression model to represent the dynamical relationships of observation symbols with time variance. The model can improve the weaknesses of standard HMM and nonstationary HMM.