Complexity of acoustic-production-based models of speech perception

Abstract

Discriminant analysis models trained on acoustic vowel production data have been found to have significant correlation with listeners' perception. Two regularised discriminant analysis models were trained on monolingual speakers' vowels. One model was trained on North Central Peninsular Spanish vowel tokens, and the other on Western Canadian English vowel tokens. For each language the model which resulted in the lowest cross-validated classification error rate was close to the least complex model possible, i.e., close to linear discriminant analysis using the variances of each acoustic variable but not the covariances between variables. In order to make cross-language vowel perception predictions the Spanish model was used to classify English vowel tokens and the English model was used to classify Spanish vowel tokens. Results suggest that monolingual North Central Peninsular Spanish listeners would assimilate most tokens of Western Canadian English /i/ and /I/ to Spanish /i/ and /e/ respectively, and thus for this combination of dialects, Spanish-speaking learners of English would not be expected to have difficulty with the English /i/-/I/ contrast.