A Fast Acoustic Model Based on Multi-Scale Feature Fusion Module for Text-To-Speech Synthesis

In the end-to-end Text-to-speech synthesis, the ability of acoustic model has important effects on the quality of the speech generated. In the acoustic model, the encoder and decoder are critical components, and usually Transformer is used. The previous works have a lack of ability to model the essential features of speech signal, as they model fixed length of features. There is also limitation of slow inference speed of the acoustic model due to the characteristics of the transformer including the high-computational multi-head self-attention layer. This limits the application of TTS in the low performance of devices such as embedded devices or mobile phones. In this paper, we propose a novel acoustic model to model the different length of features and improve the speed of generating synthetic speech and naturalness as compared to the conventional Transformer structure. Through the experiment, we confirmed that the proposed method improves the naturalness of synthetic speech and operation speed in the low performance of devices.