First International Conference on Spoken Language Processing (ICSLP 90)
This paper describes a Japanese phonetic typewriter based on HMM phone units and syllable trigrams. Even though HMM methods have considerable ability to recognize speech, it is still difficult to recognize individual phones in continuous speech without lexical information. This paper reports on a phonetic typewriter to improve HMM phone recognition performance by incorporating syllable trigrams. HMM phone units are trained using an isolated word database, and their duration parameters are modified in relation to the speaking rate. The syllable trigram tables are made from a large text database of over 35,000 syllables. Phone sequence probabilities calculated from the trigrams are combined with HMM probabilities. Limiting the number of intermediate candidates using these probabilities leads to an accurate phonetic typewriter system without excessive computation time.
Bibliographic reference. Kawabata, Takeshi / Hanazawa, Toshiyuki / Itoh, Katsunobu / Shikano, Kiyohiro (1990): "Japanese phonetic typewriter using HMM phone units and syllable trigrams", In ICSLP-1990, 717-720.