The USTC System for Blizzard Challenge 2017

This paper introduces the details of the speech synthesis system developed by the USTC team for Blizzard Challenge 2017. A 6.5-hour corpus of highly expressive children’s audiobook was released to the participants this year. A parametric system that modeling speech waveforms was built for the task. Firstly, long short term memory (LSTM)-based recurrent neural networks (RNN) were adopted for the baseline system, including tone and breaking indices (ToBI) prediction, duration modeling and acoustic modeling. Then, we proposed a generative adversarial network (GAN) based post-filtering to relieve the oversmoothing in acoustic modeling and compensate for the differences between natural and synthetic spectrum in the baseline system. At last, a WaveNet based neural vocoder was utilized to model speech waveforms from acoustic feature instead of melcepstrum vocoder. The evaluation results show the effectiveness of the submitted system.