An Isarn dialect HMM-based text-to-speech system

Abstract

This paper presents a statistical parametric text-to-speech system for the Isarn language, which is a regional dialect of Thai. The features of speech, which consist of Mel-cepstrum and fundamental frequencies, were modelled by the Hidden Markov Model (HMM). Synthetic speech is generated by converting the input text to context-dependent phonemes. Speech parameters are generated from the trained HMM models, according to the context-dependent phonemes. The parameters produced are then synthesized through a speech vocoder. In order to evaluate the intelligibility and naturalness of the proposed system, we conducted a listening test with 20 native speakers. The results indicated a mean opinion score (MOS) of the proposed system of 3.49. The word error rates (WER) within the unpredictable and predictable sentences of the proposed system were 4.28% and 0.84%, respectively.