First International Conference on Spoken Language Processing (ICSLP 90)
In this paper, we investigate training algorithms of hidden Markov model(HMM) parameters for speech recognition. Here the speech recognition problem is thought of as multicategory hypotheses testing based on the HMM method. The criterion used in a training algorithm is classified into two classes; single and multicategory criterion. We study the maximum likelihood estimation(MLE) and unification estimation(UE) methods which use single category criterion, and the maximum mutual information estimation(MMIE) and corrective estimation(CE) methods which use multicategory criterion. Also, we propose a hybrid method of training and a method of variable weighting to each category. In addition, we compare the performances of these training methods by a confusable-phoneme recognition experiment.
Bibliographic reference. Kim, N. S. / Un, Chong Kwan (1990): "Generalized training of hidden Markov model parameters for speech recognition", In ICSLP-1990, 225-228.