Experimental study of the HMMs effect on the word recognition performance

A standard approach to automatic speech recognition uses HMM whose state dependent distributions are Gaussian mixtures models. In this paper we evaluate experimentally on the automatic word recognition performance, the effect of different hidden Markov models (HMM) by varying the number of state and the number of Gaussian mixture per state. We evaluate the different models with different coding techniques: linear predictive cepstral coefficients, Mel frequency cepstral and perceptual linear predictive coefficients combined with the first derivate coefficient known as the delta coefficients, in aim to built a reference word recognition system. The system is performed using the HTK 3.1 toolkit.