Animal and computational models of development of graded vowel categories

Abstract

Some acoustic instances of phonetic segments are phonemically more compelling than others, and studies involving human infants and monkeys suggest that experience plays a critical role in modifying the manner in which subjects respond to between and within phonetic category differences. In this study, eight European starlings (Sturnusvulgaris) were trained to discriminate vowel tokens drawn from stylized distributions either of the Engish vowel categories /i/ and /I/, or of the Swedish vowel categories /y/ and /I/. Following training, responses to novel stimuli drawn from these distributions indicated that starlings’ responses generalized with facility to novel category examples. Multiple linear regression analyses revealed that responses could be well predicted on the bases of F1 and F2 values and distributional characteristics of experienced vowel sounds. Responses manifested a graded structure about the central ‘‘prototypical’’ vowel of the training category. Starling response rates (/i-I/birds) corresponded closely to adult human judgments of ‘‘goodness’’ for the English vowel sounds. Finally, a simple linear association network model trained with vowels drawn from the training set captured 95% of the variance in birds’ response rates for novel vowel tokens. [Work supported by NIDCD and NSF.]