Connectionist approach for emission probability estimation in Malayalam continuous speech recognition

Abstract

Automatic speech recognition is one active research area which can exploit the pattern recognition capabilities of artificial neural networks. Several researchers have shown that the outputs of artificial neural networks trained in multi-class classification mode can be interpreted as estimates of a posteriori probabilities of output classes. These probabilities can be used by the state-of-theart hidden Markov model for speech recognition in estimating the emission probabilities of the states of the hidden Markov model. In this paper, we explore a pairwise neural network system as an alternative approach to multi-class neural network systems to estimate the emission probabilities of the states of a hidden Markov model. Through experimental analysis it is shown that the pairwise recognition system outperforms the multiclass recognition system in terms of the recognition accuracy of spoken sentences.