Explicit pattern recognition models for speech perception

Abstract

Optimal statistical classification of arbitrary input signals can be obtained, in principle, via a Bayesian classifier, given (perfect) knowledge of the distributions of signal properties for the set of target categories. At least for certain constrained problems, such as the perception of isolated vowels, simple (imperfect) statistical pattern recognition techniques can accurately predict human listeners’ performance. This paper sketches several relatively successful case studies of the application of static pattern recognition techniques to speech perception. (Static techniques require inputs of a fixed length, e.g., F1 and F2 for isolated vowels.) Real speech clearly requires dynamic pattern recognition, allowing inputs of arbitrary length. Certain such methods, such as dynamic programming and hidden Markov models, have been widely exploited in automatic speech recognition. The present paper will describe initial attempts to apply variants of such methods to the data from a perception experiment [T. Nearey and R. Smits, J. Acoust. Soc. Am. 111 (2002)] involving the perception of three (VCV) or four (VCCV) segment strings. Practical and conceptual problems in the application of such techniques to human perception will be discussed. [Work supported by SSHRC.]