An exemplar-based model of intonation perception of statements and questions in English, Cantonese, and Mandarin

Abstract

To better understand how humans can perceive intonation from speech that includes natural variability, this study investigated whether exemplar theory could account for native listeners’ categorization of sentence intonation in English, Cantonese, and Mandarin. In each language, twenty native listeners classified gated utterances of statements and echo questions produced by native speakers. Then a computational model simulated the classification of these utterances, using an exemplar-based process of categorization that relied on F0 only. The computational model correctly classified these sentences above chance without normalizing F0 by speaker. Compared to the human listeners, the model was similarly sensitive to the cross-linguistic differences in the cues for questions, but performed worse when these cues were (partly) excluded from the utterances. These results suggest that human listeners store whole intonation patterns in memory and use additional acoustic information, along with F0, to categorize new statements and questions, in accordance with exemplar theory principles.