Generation of GMM Weights by Dirichlet Distribution and Model Selection Using Information Criterion for Malayalam Speech Recognition

Abstract

Automatic Speech Recognition is a computer-driven transcription of spoken-language into human-readable text. This paper is focused on the development of an acoustic model for medium vocabulary, context independent, isolated Malayalam Speech Recognizer using Hidden Markov Model (HMM). In this work, the emission probabilities of syllables, based on HMMs are estimated from the Gaussian Mixture Model (GMM). Mel Frequency Cepstral Coefficient (MFCC) technique is used for feature extraction from the input speech. The generation of mixture weights for GMMs is done by implementing Dirichlet Distribution. The efficiency of thus generated Gaussian Mixture Model is verified with different Information Criteria namely Akaike Information Criterion, Bayes Information Criterion, Corrected AIC, Kullback Linear Information Criterion, corrected KIC and Approximated KIC (KICc, AKICc). The accuracy of medium vocabulary, speaker dependent and isolated Malayalam speech corpus for a single Gaussian is 90.91% and Word Error Rate (WER) is 11.9%. The word accuracy and WER of the system are calculated based on the experiments conducted for multivariate Gaussians. For Gaussian mixture five, a better word accuracy of 95.24% along with a WER of 4.76% is attained and the same is verified using Information Criteria.