The Tencent speech synthesis system for Blizzard Challenge 2020

Abstract

This paper presents the Tencent speech synthesis system for Blizzard Challenge 2020. The corpus released to the partici-pants this year included a TV’s news broadcasting corpus with a length around 8 hours by a Chinese male host (2020-MH1 task), and a Shanghaiese speech corpus with a length around 6 hours (2020-SS1 task). We built a DurIAN-based speech synthesis system for 2020-MH1 task and Tacotron-based system for 2020-SS1 task. For 2020-MH1 task, firstly, a multi-speaker DurIAN-based acoustic model was trained based on linguistic feature to predict mel spectrograms. Then the model was fine-tuned on only the corpus provided. For 2020-SS1 task, instead of training based on hard-aligned phone boundaries, a Tacotron-like end-to-end system is applied to learn the mappings between phonemes and mel spectrograms. Finally, a modified version of WaveRNN model conditioning on the predicted mel spectrograms is trained to generate speech waveform. Our team is identified as L and the evaluation results shows our systems perform very well in various tests. Especially, we took the first place in the overall speech intelligibility test.