Probabilistic Modeling of Tone Perception: Autosegmental "targets" are insufficient

Deepti Ramadoss


Existing tone perception models simply match f0-value strings of stimuli to category templates. Here we propose probabilistic models of Thai tone perception motivated by phonological theory: tone categories are defined by High(H) and Low(L) target combinations associated with syllable moras. Category specific prior probability distributions are computed over these target (f0-value) sets; then, the most likely category of any given stimulus is computed. Probabilities computed by Model 1 are over targets (accuracy: 36.36%); Model 2 are over slopes between targets (accuracy: 63.6%). Overall, this research shows that phonological representations of a single cue are insufficient to account for tone perception.

Full Text: