A New Method for the Estimation of Hidden Markov Model Parameters
隐Markov模型参数估计的一种新方法

The authors present a new method for estimating the parameter values of hidden Markov models in speech processing. In stead of relying on the traditional maximum likelihood rule, this method is based on the rule of maximum recognition accuracy when the model is used directly as recognizer. The performance of this method is better than maximum likelihood estimation (MLE). Experiment has shown that the accuracy of speech recognition is about 5% higher than that of MLE.