Improving generative adversarial networks for speech enhancement through regularization of latent representations

Fan Yang, Ziteng Wang, Junfeng Li, Risheng Xia, Yonghong Yan

Abstract: Highlights•We propose a new network structure and a new loss function, which is advantageous to our model in speech enhancement under low signal to noise (SNR) environments and low resource environments.•Different from most network structures, t he new network enables us to make full use of the information carried by the clean speech signals.•The new loss allows us to obtain a more accurate speech feature re presentation from a noisy speech signal and improves the optimization direction of the network.•We explain the reasons for the excellent performance of the proposed model.•Extensive experiments demonstrate the generality of our model in a variety of speech enhancement cases.