Use of Computer Speech Understanding in Training: A Preliminary Investigation of a Limited Continuous Speech Recognition Capability.

摘要

This report describes the requirements for connected word speech recognition in training systems design. A review of the published literature and commercially available systems demonstrates that no existing capability exists which fully satisfies the unique features of speech recognition in training devices. A new approach toward limited continuous speech recognition was investigated based on examining the speech data itself to find characteristic sound structures; the order in which these sounds occur; and the time duration of, and between, these characteristic sounds. Algorithms were developed and exercised over speech data generated by a commercial preprocessor for a large number of utterances spoken by a single speaker. Characteristic structure was found for the 10 digits and the word point. Borrowing from the mathematical theory of formal languages, these sounds are termed transition letter sets. The residual data were formed into loop letter sets which are used to reduce false recognitions. Nondeterministic finite transducers, defined by the transition and loop letter sets for each vocabulary item, were exercised over entire utterances to validate the basic concepts and extract time duration data.