HMM based duration control for singing TTS

Abstract

In order to develop a HMM based singing TTS system, we need a huge singing voice database to train HMM model parameters. However there is no singing voice database publically available and the construction of it is much more difficult than that of speech database. In this paper we propose a new method to improve the naturalness of singing TTS system using HMM models from speech database. Duration control model based on the syllabic analysis is applied to adapt speech duration model to singing duration model. The proposed method results in better singing voice quality compared to the maximum likelihood generation of durations using the speech database.