Mark A. Randolph
AT&T Bell Labs., Rm. 2D-557, 600 Mountain Ave., Murray Hill, NJ 07974
A major problem in articulatory-based speech recognition is the need for the recognizer to transform an input sequence of acoustic observations into a set of parameters describing the behavior of the articulatory system during utterance production. This acoustic-to-articulatory transformation is nonunique. In addition, an economical computational framework is needed to describe and recognize patterns of articulatory behavior obtained from acoustic analysis. In this paper, a speech analysis system based on a generative model of articulatory behavior is presented. Articulatory codebooks represent the acoustic-to-articulatory relation. During speech analysis ambiguity is effectively resolved with the aid of a finite-state grammar of articulatory behavior. The grammar, in the form of a finite-state acceptor, is used to constrain codebook search. This analysis procedure is part of a word recognition system, the finite-state grammar of articulatory behavior can be derived automatically from an analysis of phonetic text as part of the process of building a recognition lexicon. This derivation procedure, as well as the speech analysis technique, is illustrated.