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

Banff, Alberta, Canada
October 13-16, 1992

Speaker Independent Word Recognition Using Continuous Matching of Parameters in Time-Spectral Form Based on Statistical Measure

Tatsuya Kimura, Mitsuru Endo, Shoji Hiraoka, Katsuyuki Niyada

Matsushita Research Institute Tokyo, Inc., Kawasaki, Japan

This paper describes a new speaker-independent speech recognition method, which effectively uses dynamic features of speech. The method uses a segment-based parameter which consists of a series of LPC cepstrum coefficients obtained during several frames, which involves the dynamic features. First, a segment-based matching based on a statistical distance measure is performed. To reduce the amount of calculation, the method utilizes a linear discriminant function for segment-level matching. Second, word-level matching is performed by accumulating segment-based likelihoods using either the DTff or the HMM. Experiments to recognize 100 Japanese city names uttered by 50 people show the validity of the present method.

Full Paper

Bibliographic reference.  Kimura, Tatsuya / Endo, Mitsuru / Hiraoka, Shoji / Niyada, Katsuyuki (1992): "Speaker independent word recognition using continuous matching of parameters in time-spectral form based on statistical measure", In ICSLP-1992, 169-172.