Computational Modelling of Tone Perception Based on Direct Processing of f0 Contours

Abstract

It has been widely assumed that in speech perception it is imperative to first detect a set of distinctive properties or features and then use them to recognize phonetic units like consonants, vowels, and tones. Those features can be auditory cues or articulatory gestures, or a combination of both. There have been no clear demonstrations of how exactly such a two-phase process would work in the perception of continuous speech, however. Here we used computational modelling to explore whether it is possible to recognize phonetic categories from syllable-sized continuous acoustic signals of connected speech without intermediate featural representations. We used Support Vector Machine (SVM) and Self-organizing Map (SOM) to simulate tone perception in Mandarin, by either directly processing f0 trajectories, or extracting various tonal features. The results show that direct tone recognition not only yields better performance than any of the feature extraction schemes, but also requires less computational power. These results suggest that prior extraction of features is unlikely the operational mechanism of speech perception.