Prosodic Aids to Speech Recognition:Acoustic-Prosodic Patterns in Selected English Phrase Structures.

摘要

The final two studies have been completed in a four year effort on developing prosodic aids to speech recognition. A procedure for using intonational phrase boundaries to select among alternative word and phrase hypotheses has been developed, refined, and tested by hand analyses of sixteen sentences. This procedure was designed for use with the BBN HWIM speech understanding system, and was totally implemented, but not tested before the end of the BBN contract with ARPA. Comparisons of control and parsing traces with acoustically detected phrase boundaries did show, however, that intonational boundaries could help select correct words and phrase structures and avoid erroneous hypotheses. An experimental study of acoustic prosodic patterns in 255 sentences showed several useful prosodic regularities. Over 91% of the syllables were correctly located, and 92% of the stressed syllables were correctly categorized as stressed, while 76% of the syntactic phrase boundaries were detected. Exactly which phrases are or are not preceded by intonational phrase boundaries was determined. Intonation contours were very firmly shown to involve rising pitch until the first stress, progressively lower pitch in succeeding stresses, and a terminal fall (for declaratives, commands, and WH questions) or rise (for yes/no questions). Parentheticals were clearly marked by disjunctures, large Fo variations, and other prosodic features. Contrastive phrase structures could be detected from prosodic cues.