A discriminative hidden Markov model recognizer with neural network postprocessor

This paper is concerned with the problem of improving recognition accuracy of hidden Markov models (HMM) for sequential pattern recognition. It is argued that maximum-likelihood estimation of the HMM parameters via the forward-backward algorithm may not lead to values which maximize recognition accuracy. We introduce a hybrid method with neural network postprocessor which is aimed at minimizing the number of recognition errors. This method exploits the discrimination capability of neural network classifier while using HMM formalism to capture the dynamics of input patterns. Although it has not been proved that the presented method is a kind of maximum mutual information estimation, experimental results with online handwriting characters suggest that it leads to fewer recognition errors than can be obtained with the conventional recognition method