Myanmar Text-to-Speech Synthesis Using End-to-End Model

Abstract

In this paper, we propose a Myanmar speech synthesis system based on an End-to-End neural network model, which integrates the Myanmar phone model into the Tacotron2 End-to-End model. Based on the Seq2seq model architecture, we use phone-level embedding to form a feature prediction network from phone sequences to Mel spectrum, and combine with a semi-supervised speech generation network to generate high-quality Myanmar synthesized speech. In addition, we introduced the BERT pre-training decoder module to assist the phone feature extraction, which reduces the system's dependence on the phone feature extraction network and improve the text feature richness. Compared with other Myanmar speech synthesis systems, this method effectively improves the naturalness and accuracy of synthesized speech under low resource conditions.