eduzhai > Applied Sciences > Engineering >

Dynamic Attention Based Generative Adversarial Network with Phase Post-Processing for Speech Enhancement

  • Save

... pages left unread,continue reading

Document pages: 6 pages

Abstract: The generative adversarial networks (GANs) have facilitated the developmentof speech enhancement recently. Nevertheless, the performance advantage isstill limited when compared with state-of-the-art models. In this paper, wepropose a powerful Dynamic Attention Recursive GAN called DARGAN for noisereduction in the time-frequency domain. Different from previous works, we haveseveral innovations. First, recursive learning, an iterative training protocol,is used in the generator, which consists of multiple steps. By reusing thenetwork in each step, the noise components are progressively reduced in astep-wise manner. Second, the dynamic attention mechanism is deployed, whichhelps to re-adjust the feature distribution in the noise reduction module.Third, we exploit the deep Griffin-Lim algorithm as the module for phasepostprocessing, which facilitates further improvement in speech quality.Experimental results on Voice Bank corpus show that the proposed GAN achievesstate-of-the-art performance than previous GAN- and non-GAN-based models

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×