Hidden Markov models (HMMs) are widely used for automatic speech recognition because they have a powerful algorithm used in estimating the models' parameters, and achieve a high performance. Once a structure of the model is given, the model's parameters are obtained automatically by feeding training data. There is, however, no effective design method leading to an optimal structure of HhlMs. In this paper, we propose a new application of a genetic algorithm to search out such an optimal structure. In this method, the left-right structures are adopted for HMMs and the likelihood is used for the fitness of the genetic algorithm. We report the results of our experiment showing the effectiveness of the genetic algorithm in automatic speech recognition.

No abstract

No abstract

SUMMARYAn automatic speech recognition method is proposed using the Markov model in which a large number of states is adopted to model the transitional characteristics of speech more accurately. Unlike the traditional HMM, the feature vector of this proposed model is considered to be the parameter of the state of the Markov model. First, the transition-probability of the state and the symbol output probability are estimated on the assumption that they are represented by multidimensional normal density functions of the feature vector. The optimal time sequence is determined by using DPmatching.Next, the multistate model is obtained by quantizing the feature vector space and calculating (or sampling) the value of the probability density function at each code vector (at multipoints). The probability distribution of the multistate Markov model is adapted by the retraining method to be more accurate in distribution, The resulting recognizer is evaluated on a vocabulary of English four-digit numerals. The multistate model has attained a recognition a r e of 98.2 percent, i.e., 1.6 percent higher than that of a five-state traditional HMM. The processing time for the reference pattern generation is one-fiftieth that of the HMM.

No abstract