Reducing over-smoothness in HMM-based speech synthesis using exemplar-based voice conversion

Abstract

Speech synthesis has been applied in many kinds of practical applications. Currently, state-of-the-art speech synthesis uses statistical methods based on hidden Markov model (HMM). Speech synthesized by statistical methods can be considered over-smooth caused by the averaging in statistical processing. In the literature, there have been many studies attempting to solve over-smoothness in speech synthesized by an HMM. However, they are still limited. In this paper, a hybrid synthesis between HMM and exemplar-based voice conversion has been proposed. The experimental results show that the proposed method outperforms state-of-the-art HMM synthesis using global variance.