First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Phoneme Recognition by Combining Bayesian Linear Discriminations of Selected Pairs of Classes

Tatsuya Kawahara, Toru Ogawa, Shigeyoshi Kitazawa, Shuji Doshita

Department of Information Science, Kyoto University, Kyoto, Japan

A new phoneme recognition method based on Bayesian linear discrimination (BLD) is presented. The conventional BLD lowers the performance as the number of classes to be discriminated becomes larger. To overcome this defect, we propose the pair-wise discrimination method which combines BLDs of the pairs of the classes. A given sample is recognized as the class which is supported or is not denied by most pairs. This method realizes high accuracy recognition but needs much computation and storage. Therefore, an algorithm to select effective pairs for overall discrimination is discussed. To measure the effectiveness of each pair, we count the occurrence of the first candidate and the second candidate combination pairs for all the samples and eliminate those pairs whose occurrence is rare. Thus we could reduce the necessary pairs for discrimination to about one fourth within 1% accuracy decrease.

Full Paper

Bibliographic reference.  Kawahara, Tatsuya / Ogawa, Toru / Kitazawa, Shigeyoshi / Doshita, Shuji (1990): "Phoneme recognition by combining Bayesian linear discriminations of selected pairs of classes", In ICSLP-1990, 229-232.