First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Syllable Structure Parsing for Continuous Speech Recognition

Shigeru Ono

C & C Information Technologmy Research Laboratories, NEC Corporation, Kawasaki, Japan

This paper describes a scheme to deal with allophonic and coarticulatory variations for phoneme-based continuous speech recognition and a probabilistic algorithm to parse syllable structure from acoustic speech realizations. In the scheme, phonological objects are represented in terms of "syllable features "-syllable positions- and "phoneme features" - distinctive features-, and they are organized within hierarchical structures. The constituent features of the structures are associated with the acoustic realizations through probabilistic measure. In the algorithm, syllable structure is parsed from the acoustic realizations by applying the acoustic-phonological constraints and the collocational restrictions involved in the internal constituent features. Performance results for 15 test sentences spoken by 5 male speakers that phonemes are recognized at 90.5% accuracy, and syllable structure is parsed at 79.7% accuracy.

Full Paper

Bibliographic reference.  Ono, Shigeru (1990): "Syllable structure parsing for continuous speech recognition", In ICSLP-1990, 269-272.