Using gaussian mixtures on triphone acoustic modelling-based punjabi continuous speech recognition

Abstract

Continuous speech recognition for a particular language is always an area which relies, for its performance, on these major aspects: acoustic modelling and language modelling. Gaussian mixture model-hidden Markov model (GMM–HMM) is a part of acoustic modelling. These components are applied at the back end of ASR design to accurately and efficiently convert continuous speech signal to corresponding text. Triphone-based acoustic modelling makes use of two different context-dependent triphone models: word-internal and cross-word models. In spite of active research in the field of automatic speech recognition for a number of Indian and foreign languages, only few attempts have been made for Punjabi language, specially, in the area of continuous speech recognition. This research paper is aimed at analysing the impact of GMM–HMM-based acoustic model on the Punjabi speaker-independent continuous speech recognition. Recognition accuracy has been determined at word and sentence levels, respectively, with PLP and MFCC features by varying Gaussian mixtures from 2 to 32.