Speech Enhancement by CycleGAN Using Feature Map Regularization

  • Branislav Popović Faculty of Technical Sciences, University of Novi Sad


Highly promising speech enhancement results are recently obtained using an unsupervised CycleGAN approach, comparable to paired dataset neural network approach. However, very often, only a limited amount of noisy speech data is available. Therefore, a semi-supervised CycleGAN approach has been proposed, relying on augmented data samples. Another feature map regularized CycleGAN approach has also been proposed and applied in an image-style translation task, obtaining significant improvements on several standard databases. The feature map regularized CycleGAN approach is combined with the aforementioned semi-supervised learning approach and applied within a speech enhancement task. Significant improvements are obtained in terms of several standard measures using the proposed algorithm in comparison with the baseline algorithm as well as the augmented CycleGAN approach.