Design of Discriminators in GAN-Based Unsupervised Learning of Neural Post-Processors for Suppressing Localized Spectral Distortion

We present a discriminator that focuses on local features of spectrograms for a generative adversarial network (GAN)-based training of a post-processor that reduces distortions in enhanced signals from nonlinear speech enhancement (SE). In learning GANs, the design of discriminators is one of the essential points. Several prior works demonstrated the effectiveness of the multi-scale discriminator in speech enhancement, which exploits the outputs of multiple layers with different granularities and considers both local and global features. However, we confirmed that distortions caused by nonlinear SE are more localized than noise in microphone observations, which implies that it may be better to focus more on local features for distortion reduction of enhanced signals. Therefore, we construct a discriminator that considers such properties and examine its effectiveness in the CycleGAN-based unsupervised learning. In experiments, we generated enhanced speeches with a nonlinear area-enhancement approach and explored the optimal receptive field for the discriminator. Experimental comparisons demonstrated that the discriminator that focused more on locality gave a higher performance.