Chinese Speech Synthesis System Based on End to End

摘要

This paper describes an end-to-end Chinese speech synthesis scheme, a neural network architecture for speech synthesis directly from text, which is based on the improvement of tacotron2. In order to improve the original seq2seq model, multi attention mechanism is used to replace the original position-based attention mechanism to improve the sound quality of synthetic speech. The prosodic style coding module is added to better show the prosodic and stylistic characteristics of the synthesized speech. Then an LpcNet vocoder is used to replace the original WaveNet to generate time-domain waveform, which improves the synthesis efficiency and reduces the synthesis time. The experiments show that the speech synthesis system is effective and timely. The synthesized speech is natural and has good prosodic style.