Research on underground speech enhancement technology based on generative adversarial network

Abstract

Because of the problems such as speech interaction and speech input difficulty caused by high noise intensity and unclear sources in underground mine, this paper proposes a speech enhancement system based on a Time-domain Generative Adversarial network, which is used in the front end of underground communication or speech recognition to improve the quality of speech information transmission and improve work efficiency. Aiming at the problems of ignoring the feature information between channels and training instability when extracting speech information in a time domain generative adversarial network, this paper introduces channel attention and Relativistic Average Generative Adversarial Network to optimize. The experimental results show that compared with other models, the proposed model can more effectively remove the downhole noise.