Perceptual learning of vowels in a neuromorphic system

Abstract

Linear neuromorphic systems assume that neurally encoded stimuli are mapped through linear combinations of a set of underlying features (represented as the eigenvectors of the matrix of learned associations) onto perceptual categories. Learning in this type of system may be through an auto-associative process that induces prototypes that represent the central tendency of each perceptual category. It has been proposed that human listeners learn vowel categories through just this type of computational mechanism. To investigate this claim, we trained an auto-associative network on a set of American English vowels averaged over talkers. We examined performance of this network for learning and classifying vowels produced by the average of tokens from several male talkers. In addition, we compared the effects of different auditory coding representations on recognition performance for the male vowels. Since, for these models it is necessary for the stimulus set to be linearly independent, the perceptual representation of the vowels can affect learning. Furthermore, we investigated the extent to which the underlying feature space learned from a set of vowels is shared between talkers. For this type of network to be a plausible model of vowel perception, it must be capable of perceptual constancy across talkers. Finally, we compared the effects of different learning algorithms on the development of vowel categories in perceptual space, and investigated the ability of a paired-associate model to develop vowel categories. We will discuss the results of these studies and their implications for simple, network models of speech perception and talker normalization.