Speaker Identification System Using Hidden Markov Model Based on Local Distribution Membership

Abstract

The work described in this paper addresses an application of Hidden Markov Model (HMM) with modification in observation probability using two approaches for membership values, i.e. Euclidean distance base and kernel function base. Mel-Frequency Cepstrum Coefficients (MFCC) used as feature extraction. In this research we use "pudesha'' as a keyword to identify each speaker and modeled by left-right HMM which its states are clustered using Fuzzy C-Mean clustering. Each new observation, we compute the membership value to every cluster and choose the biggest one as the membership value of the observation to appropriate hidden state. In the unguided-utterance and limited training data, experimental results show that our methods recognize better than classical HMM that uses Normal Distribution as observation probability. It is also showed that the use of Normal distribution for observation probability leads to a singularity problem in computing the inverse of covariance matrix, especially for limited training data. In our proposed approaches, the singularity problem will not occurs, since we do not need to compute the inverse of covariance matrix.