HMM speech recognition with reduced training

Abstract

One of the problems faced in automatic speech recognition is the amount of training required to adapt the machine to the speaker way of pronunciation. To a certain extent, the accuracy of correct recognition is proportional to the amount of training and adaptation carried out. This is especially true when a large vocabulary is involved. For certain applications, it is desirable that the training requirement be reduced to the bare minimum without sacrificing the accuracy of recognition. The minimum number of training required to achieve an acceptable degree of accuracy for a speaker dependent speech recognition system based on the hidden Markov model (HMM) is investigated. A method is also proposed which retains the same degree of accuracy of recognition with much reduced training.