Speaker-Independent Consonant Recognition by Integrating Discriminant Analysis and HMM

Abstract

In this paper, we propose a new consonant recognition method which integrates two stochastic method : discriminant analysis and HMM (Hidden Markov Models). Discriminant Analysis is effective to analyze local patterns around the reference-point of a consonant such as a burst point. This method, however, is based on the assumption that the reference-point is detected precisely. HMM is able to extract the global dynamic features of a consonant from the preceding vowel to the following vowel and needs no explicit segmentation of speech. But it is hard to discriminate between similar consonants with HMM due to the quantization of input pattern vectors. Our new method constructs HMM with discriminant analysis front-end and recognizes consonants by combining the score obtained by discriminant analysis and the score by HMM. In recognition experiments of all the Japanese consonants in mono-syllables, this integrated method achieved the recognition rate of 92.1 %, which is higher by 5~15 % than the case using either of two methods alone.