Word Verification in a Speech Understanding System.

A word verification component of a speech understanding system is intended to give a reliable yesno decision as to whether a hypothesized word is present at a particular region of the acoustic input. A verification stage is a desirable system component because spectrogram-reading experiments have shown that preliminary acoustic-phonetic analyses are inherently ambiguous, whereas all the detailed acoustic cues must be consistent with a particular work hypothesis or the hypothesis can be rejected in a straight-forward way. The author proposes a word verification strategy in which a synthesis by rule program is used to predict the acoustic characteristics of a given word. This analysis-by-synthesis paradigm facilitates the statement of what are essentially generative phonological and phonetic rules. Spectral matching criteria and dynamic time normalization procedures are proposed for determining whether the synthesis matches the acoustic input.