Tone recognition of Thai continuous speech using Fujisaki's model

This paper proposes a recognition algorithm, using Fujisaki's model parameters as features for tone recognition of a predominantly tonal language, without limitation of the number of syllables in the utterance. The algorithm is composed of 4 steps. Initially, the F0 speech contour is extracted. Fujisaki's model parameters of the whole utterance are extracted manually. The utterance-base tone commands are then segmented into syllable-base tone commands and are used as classification features and fed into an MLP neural network recognition system. The experimental results show that the proposed system, when Thai continuous speech was selected for test, achieves an average recognition rate of 96.35%.