The Ajmide Text-To-Speech System for Blizzard Challenge 2020

Abstract

This paper presents the Ajmide team’s text-to-speech system for the task MH1 of Blizzard Challenge 2020. The task is to build a voice from about 9.5 hours of speech from a male native speaker of Mandarin. We built a speech synthesis system in an end-to-end style. The system consists of a BERT-based text front end that process both Chinese and English texts, a multi-speaker Tacotron2 model that converts the phoneme and linguistic feature sequence into mel spectrogram, and a modified WaveRNN vocoder that generate the audio waveform from the mel spectrogram. The listening evaluation results show that our system, identified by P, performs well in terms of naturalness, intelligibility and the aspects of intonation, emotion and listening effort.