Critique: The potential role of speech production models in automatic speech recognition

Abstract

This paper presents a critique (by a ‘‘speech technologist’’) of ‘‘The potential role of speech production models in automatic speech recognition’’ by R. C. Rose, J. Schroeter, and M. M. Sondhi which was presented at the 1994 ASA Special Session on ‘‘Speech Recognition and Perception from an Articulatory Point of View.’’ While in general agreement with the points raised in the Rose paper, this article focuses on issues such as (a) whether articulatory information should be ‘‘recovered’’ as part of a recognition process as opposed to that process being ‘‘constrained’’ by articulatory information, (b) how optimum recognition performance is achieved by compromising the ‘‘quality’’ of the underlying generative model, and (c) if recent developments in speech pattern modeling will lead to more powerful generative models of speech. Finally, it is noted that the special session highlighted the fact that a convergence is long overdue between the speech science and speech technology communities toward common (mathematically and computationally based) theories of ‘‘speech pattern processing.’’