MINIMUM CLASSIFICATION ERROR RATE METHODS FOR SPEECH RECOGNITION

A critical component in the pattern matching approach to speech recogn ition is the training algorithm, which aims at producing typical (refe rence) patterns or models for accurate pattern comparison, In this pap er, we discuss the issue of speech recognizer training from a broad pe rspective with root in the classical Bayes decision theory, We differe ntiate the method of classifier design by way of distribution estimati on and the discriminative method of minimizing classification error ra te based on the fact that in many realistic applications, such as spee ch recognition, the real signal distribution form is rarely known prec isely, We argue that traditional methods relying on distribution estim ation are suboptimal when the assumed distribution form is not the tru e one, and that ''optimality'' in distribution estimation does not aut omatically translate into ''optimality'' in classifier design, We comp are the two different methods in the context of hidden Markov modeling for speech recognition, We show the superiority of the minimum classi fication error (MCE) method over the distribution estimation method by providing the results of several key speech recognition experiments, In general, the MCE method provides a significant reduction of recogni tion error rate.