Speech Recognition using Hidden Markov Model with Neural Network as Probability Estimators

Rachit Kr. Rastogi,

Department of Computer Engg., College Of Technology,

Govind Ballabh Pant University of Agriculture & Technology,.

Pantnagar, Uttaranchal, INDIA -263145

Abstract

Automatic speech recognition, an exciting area of research, can allow us to interact with computers using ‘spoken commands’. Hidden Markov Model (HMM) is the most widely used statistical speech recognition technique. In this paper, I describe HMM based speech recognition system and point to it’s limitations, some of which can be alleviated by use of Neural Networks. I describe the use of Multilayer Perceptron (MLP) as probability estimators in HMM framework The practical implementation of this hybrid HMM/MLP based speech recognition system is also illustrated. Experimental results on isolated word, limited vocabulary task are also reported. Particularly, the effect of learning factor variation is training shows that by tuning it properly, the recognition performance can be improved significantly.