IMPLEMENTATION OF GRADIENT-BASED ESTIMATION OF HMM PARAMETERS

IntroductionIn Hidden Markov Model ( HMM) - basedspeech recognition systems,it is very importanttoset the parameters of HMMs correctly.Conven-tional training algorithms such as Baum- Welch al-gorithm,are based on the criterion of maximumlikelihood( ML) .The strength of Baum- Welch al-gorithm lies in its fast convergence andmonotonous improvement1] . For other criteriasuch as maximum mutual information,however,such an algorithm does not exist.In this case,acommon resort is a gradient- b…

IMPLEMENTATION OF GRADIENT-BASED ESTIMATION OF HMM PARAMETERS

The most popular estimation method for HMMs is Baum Welch algorithm, which is based on the maximum likelihood(ML) criterion. For other criteria, such as Maximum Mutual Information (MMI) criterion, such an algorithm does not exist. In this case, a gradient based method is considered. With the complexity of the objective function, the computation of the gradients has to be solved before it can be applied to this problem. This paper proposed an implementation method of the gradient based method. Experimental results indicate that this method produces monotonous improvement like Baum Welch algorithm.