Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Automatic Speech Recognition Using Dynamically Defined Speech Units

L. Deng, H. Sameti

Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Ontario, Canada

Dynamically-defined, sub-phonemic units of speech are proposed and constructed by way of phonological assimilation of one or more major articulatory features. Assimilation of these features spans several phonetic segments, accounting for cross-phone context dependence in speech. A statistical theory is developed that endows these new speech units with computational and parameter-learning capability sufficiently powerful for their use as the primitive units in designing a feature-based automatic speech recognition system. The effectiveness of a feature-based speech recognition system exploiting the dynamically-defined speech units is demonstrated in preliminary evaluation experiments using a standard phonetic recognition task designed from TIMIT database.

Full Paper

Bibliographic reference.  Deng, L. / Sameti, H. (1994): "Automatic speech recognition using dynamically defined speech units", In ICSLP-1994, 2167-2170.