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

Banff, Alberta, Canada
October 13-16, 1992

Continuous Mixture HMM-LR Using the A* Algorithm for Continuous Speech Recognition

Kouichi Yamaguchi (1), Shigeki Sagayama (1), Kenji Kita (1), Frank K. Soong (2)

(1) ATR Interpreting Telephony Research Laboratories, Kyoto, Japan
(2) AT&T Bell Labs., Murray Hill, New Jersey, USA

This paper describes a new continuous mixture HMM version of the HMM-LR continuous speech recognition system. Each HMM output probability density function is characterized by a 34-component Gaussian mixture. The system performs comparably to the previous system (discrete HMM-LR) even though it does not use duration control requiring considerable computation effort. We also describe two new search methods; an A* algorithm based and a hybrid best-first search. These two new algorithms are experimentally compared with conventional beam search techniques. The A* algorithm based search has been found effective in reducing the number of phoneme verifications by half with little degradation of recognition performance.

Full Paper

Bibliographic reference.  Yamaguchi, Kouichi / Sagayama, Shigeki / Kita, Kenji / Soong, Frank K. (1992): "Continuous mixture HMM-LR using the a* algorithm for continuous speech recognition", In ICSLP-1992, 301-304.