A Prosodic Mandarin Text-to-Speech System Based on Tacotron

The Tacotron performs well in English speech synthesis and successfully aligns two arbitrary sequences from different domain in an automatic way. However, to introduce Tacotron into Mandarin Chinese Text-to-Speech (TTS), a prosody system is needed for generating more natural speech. This paper proposes a practical method to involve the prosodic annotation into Tacotron training for Mandarin Chinese synthesis system. A prosody model predicting the prosodic boundaries from the given text serves as the front-end system in our approach, followed by a Tacotron synthesis system trained with well-labeled TTS database containing the prosodic annotations. Under subjective evaluation in terms of the prosody, results show that the synthesis system performs better by adding the prosodic system as the front-end system for Tacotron.