Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Compact-Size Speaker Independent Speech Recognizer for Large Vocabulary Using "Compats" Method

Tatsuya Kimura, Hiroyasu Kuwano, Akira Ishida, Taisuke Watanabe, Shoji Hiraoka

Matsushita Research Institute Tokyo, Inc., Kawasaki, Japan

This paper describes the development of a speaker-independent speech recognizer for large vocabulary which performs well in practical applications. The method named "COMPATS(COntinu ous Matching of PArameters of Time-Spectral form)"[l] is used for recognition accuracy. A combination of the segment-based parameter and the statistically based distance measure given by the linear discriminant function make it possible to use the dynamic features of speech effectively and precisely. To realize a small sized recognizer a two-stage matching procedure is developed. In the first stage, rough but high speed matching for all templates finds a small number of appropriate candidates. For noise robustness a word-spotting technique is used in this stage. In the second stage, precise match ing is performed for the candidates obtained in the first stage and then the final recognition result is determined. Using this technique we implemented a 500-word recognizer on a small board with a single DSP chip.

Full Paper

Bibliographic reference.  Kimura, Tatsuya / Kuwano, Hiroyasu / Ishida, Akira / Watanabe, Taisuke / Hiraoka, Shoji (1994): "Compact-size speaker independent speech recognizer for large vocabulary using "compats" method", In ICSLP-1994, 1379-1382.