Comparing Human Perception to Computational Classifications of Lexical Tones

Abstract

This dissertation analyzed the tonal and acoustic properties of utterances produced from five native Thai speakers. The computational model produced classifications based on predictions made by a Hidden Markov Model that simulates tone perception and categorization. The computational model tested the categorization of stimuli taken from both citation and continuous contexts of Thai tonal data, in order to compare the performance of the computational model on both clear and naturalistic stimuli. Two perception experiments were also conducted, involving human listeners, for the purpose of comparing their behavior to that of the computational model. The results reveal that the classifications of lexical tone categories made by the computational model yield some dissimilar learning patterns to that found in human perceptual learning of the same categories.