An HMM-based Vietnamese speech synthesis system

Abstract

This paper describes an approach to the realization of a Vietnamese speech synthesis system applying a technique whereby speech is directly synthesized from Hidden Markov models (HMMs). Spectrum, pitch, and phone duration are simultaneously modeled in HMMs and their parameter distributions are clustered independently by using decision tree-based context clustering algorithms. Several contextual factors such as tone types, syllables, words, phrases, and utterances were determined and are taken into account to generate the spectrum, pitch, and state duration. The resulting system yields significant correctness for a tonal language, and a fair reproduction of the prosody.