The OPPO System for the Blizzard Challenge 2020

Abstract

This paper presents the OPPO text-to-speech system for Blizzard Challenge 2020. A statistical parametric speech synthesis based system was built with improvements in both frontend and backend. For the Mandarin task, a BERT model was used for the frontend, a Tacotron acoustic model and a WaveRNN vocoder model were used for the backend. For the Shanghainese task, the frontend was built from scratch, a Tacotron acoustic model and a MelGAN vocoder model were used for the backend. For the Mandarin task, evaluation results showed that our proposed system performed best in naturalness, and achieved near-best results in similarity. For the Shanghainese task, we got poor results in most indicators.