Cycle-consistent InfoGAN for speech enhancement in various background noises

Speech enhancement is one of the crucial research topics applied to various fields. In addition, due to the progress of wireless communication technology, the need for speech enhancement research to remove various background noise occurring in the real world is increasing. Recently, a speech enhancement model based on generative adversarial learning, which can build a significant loss function by itself, has been proposed and outperformed the conventional methods. However, these models assume parallel datasets for learning, and there is a problem that the performance decreases for the signal containing various kinds of noise. This paper proposes a novel speech enhancement model based on generative adversarial network (GAN). The proposed method additionally uses cycle-consistency loss for learning on non-parallel datasets, where the InfoGAN mechanism is used to cluster noise information in an unsupervised learning manner. The proposed model can form cluster-specific mapping by using the obtained clustering information. We quantitatively verify the speech enhancement performance of the proposed method through several metrics such as MOS, SI-SNR, and PESQ, and achieve about 55% better MOS performance than the previous GAN-based models.