Application of Pair-Wise Discrimination Method to Japanese Consonant Recognition

Abstract

We have proposed pair-wise discrimination method which discriminates multiple classes by combining the results of two-class discriminant analyses performed on the pairs of classes. In this paper, we discuss its application to recognition of all the Japanese consonants. 26 consonant classes for recognition are decided based on the phonetic category and statistical analysis by considering the influence of the following vowels. Effective pairs to discrimination are selected among all the possible 26C2 pairs. Experimental results show that, as the number of classes increases, the performance of pair-wise discrimination method is outstanding while the conventional one-stage discriminant analysis using common variables for all the classes lowers its recognition rate. It was also confirmed that minimax method is most effective in combining the results of two-class discrimination.