Thai Phoneme Soft Segmentation and Recognition Using Hidden Markov Models

In this paper, we propose a Thai phoneme recognition system with a soft phoneme segmentation. The soft phoneme segmentation technique is based on the characteristics of Thai language in that the vowel is the core of a syllable. The recognition system utilizes the discrete hidden Markov model to recognize the Thai phonemes, i.e., 21-class initial consonants, 18-class vowels, and 9-class final consonants. We use the Mel frequency with perceptual linear prediction as the features of a phoneme. We experiment the recognition system on both speaker-dependent and speaker-independent data sets recorded from 30 speakers. The experimental results show promising recognition performances in both cases.