A New Method for Improving Generative Adversarial Networks in Speech Enhancement

Abstract

Recent advances in deep learning-based speech enhancement techniques have shown promising prospects over most traditional methods. Generative adversarial networks (GANs), as a recent breakthrough in deep learning, can effectively remove additive noise embedded in speech, improving the perceptual quality [1]. In the existing methods of using GANs to achieve speech enhancement, the discriminator often regards the clean speech signal as real data and the enhanced speech signal as fake data; however, this approach may cause feedback from the discriminator to fail to provide sufficient effective information for the generator to correct its output waveform. In this paper, we propose a new method to use GANs for speech enhancement. This method, by constructing a new learning target for the discriminator, allows the generator to obtain more valuable feed-back, generating more realistic speech signals. In addition, we introduce a new objective, which requires the generator to generate data that matches the statistics of the real data. Systematic evaluations and comparisons show that the proposed method yields better performance compared with state-of-art method-s, and achieves better generalization under challenging unseen noise and signal-to-noise ratio (SNR) environments.