Improved speech enhancement algorithm based on generative adversarial networks

Abstract

According to recent research, the techniques of speech enhancement have been increasingly improved to a quite high-level, especially for speech denoising problem. We need to blindly separate the speech audio signal from background noise. This task is challenging because the speech waveform and the noise waveform are superimposed, and there are no simple features that allow the separation of the two. In this thesis, we investigate possible improvements of using generative adversarial networks to perform speech de-noising.