CycleGAN-based speech enhancement for the unpaired training data

Speech enhancement is an important task of improving speech quality in noise scenario. Many speech enhancement methods have achieved remarkable success based on the paired data. However, for many tasks, the paired training data is not available. In this paper, we present a speech enhancement method for the unpaired data based on cycle-consistent generative adversarial network (CycleGAN) that can minimize the reconstruction loss as much as possible. The proposed model employs two discriminators and two generators to preserve speech components and reduce noise so that the network could map features better for the unseen noise. In this method, the generators are used to generate the enhanced speech, and two discriminators are employed to discriminate real inputs and the outputs of the generators. The experimental results showed that the proposed method effectively improved the performance compared to traditional deep neural network (DNN) and the recent GAN-based speech enhancement methods.