A model of auditory pattern analysis based on component-relative-entropy.

Abstract

Recently results from detection and information theory have been used to predict performance of human observers in experiments involving the discriminability of changes in individual components of unfamiliar or uncertain tone patterns [R. A. Lutfi, J. Acoust. Soc. Am. 91, 3391-3401 (1992)]. In this paper, the theoretical analysis is expanded to yield a general rule of pattern analysis that accounts for nearly all of the major results of such experiments conducted over the last 18 years. The rule can be stated as follows: Component discriminability in an unfamiliar tone pattern (profile or sequence) is a linearly increasing function of the component's relative entropy (CoRE) in the pattern. Ancillary: The threshold for detection of a change decreases by one order of magnitude (factor of 10) for each one bit increase in CoRE. This precise relation of threshold to CoRE is demonstrated repeatedly in a review of past studies, and is shown to account for the relative effects of a variety of important variables and their interactions. Such variables include the total duration of the pattern; the number of tones in the pattern; the number of tones subject to change; the relative level and duration of individual tones in the pattern; the relative variability and the physical dimension of the tones' parameters that are subject to change; the number and position of targets in the pattern; the psychophysical procedure used (method of adjustment versus same-different), the type of discrimination task (frequency versus intensity discrimination), and the manner of presentation of the tones (simultaneous or sequential). The CoRE rule seems to reflect a general property of auditory analysis wherein the perception of patterns in an ensemble is dominated by those features that dominate the variance of the ensemble.