Animal models of speech sound categorization

Abstract

Several recent studies have examined phonetic category formation by training nonhuman animals (e.g., birds) to respond differentially to distributions of speech sounds. Animal models provide good tests of theories of perceptual category formation because one can completely control the input distributions and obtain a fine-grained sampling of responses to the stimulus space. In studies using vowel sound distributions, animals’ response structures reveal two salient characteristics. First, there is a tendency to respond most strongly to stimuli that are furthest from the boundaries between training distributions. Second, there is evidence that the structure of responses is a function of the statistics of the training input. This can be seen as a peak in response for stimuli near the centroid of the input distribution. These same response structure characteristics can also be seen in data collected from humans in categorization tasks. It is possible that these two characteristics are indicative of two systems of learning: one functional and one statistical. A similar dichotomy has also been recently proposed in the visual categorization literature. The implications of these results for theoretical and computational models of phonetic acquisition will be discussed. [Work supported by NSF and NIH.]