A new method for estimation of hidden Markov model parameters

Abstract

An algorithm for estimating the parameters of a hidden Markov model (HMM) is presented. In this algorithm, the rule of parameter estimation is used to maximize the recognition accuracy or to minimize the probability of error of the recognizer which is based on hidden Markov models instead of maximizing the likelihood function in maximum likelihood estimates (MLEs). The performance of minimum probability error (MPE) estimates is better than that of MLEs. Since MPE is much more complex in computation than an MLE, a simplified implementation of MPE which is much more moderate in computation is given. Experiments on speech recognition based on HMMs show that the accuracy of the recognizer trained by the estimation method has about a 5% improvement over the MLE.<>