Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Performance of Speaker-Independent Japanese Recognizer as a Function of Training Set Size and Diversity

O. Shirotsuka (1), G. Kawai (2), Michael Cohen (2), J. Bernstein (2)

(1) NTT Data Communications Systems Corporation, Kanagawa, Japan
(2) SRI International, Menlo Park, California, USA

Experiments investigated the effects of training set size and diversity of speech data in training an HMM-based, speaker-independent, continuous Japanese speech recognition system. Two different types of diversity were investigated: speaker diversity and phonetic diversity. The results indicate that greater amounts of training data improve recognition performance and that, given a fixed amount of training data, greater diversity of training materials both in terms of speakers and phonetic contexts improve recognition performance.

Full Paper

Bibliographic reference.  Shirotsuka, O. / Kawai, G. / Cohen, Michael / Bernstein, J. (1992): "Performance of speaker-independent Japanese recognizer as a function of training set size and diversity", In ICSLP-1992, 297-300.