Phoneme and tonal accent recognition for Thai speech

In this paper, we investigate the application of a phoneme recognition system with a soft phoneme segmentation procedure for Thai speech. In addition, we propose a new method to classify the tonal accent of a syllable. The recognition system classifies Thai phonemes, including the 21-class initial consonants, the 18-class vowels, and the 9-class final consonants, using discrete hidden Markov models. Two features, i.e., the Mel frequency with perceptual linear prediction and the Mel frequency cepstrum coefficients, are compared to investigate their utilities in phoneme recognition. Neural networks are applied to classify the 5-class tonal accents by using the temporal variation of pitch frequencies across syllables as features. Speaker-dependent and speaker-independent data sets recorded from 30 speakers are used to test our recognition system. The experimental results show promising recognition performances for the phonemes and tonal accents in both data sets.