Denoising Speech Signals with Hifi-Coulomb-GANs

Recorded speech signals often contain noise that affects the quality of signal and reduces intelligibility. Several studies have used Generative Adversarial Networks (GANs) to remove artifacts improve However, GANs can suffer from gradient vanishing or explosion reduce their effectiveness in denoising. To mitigate vanishing, we applied CoulombGAN architecture denoising using a model structure similar Hifi-GAN, current state art denoiser. We call this new Hifi-CoGAN. WaveNet generator denoise signals, PostNet for general cleanup, Multi-Resolution Discriminator evaluate relative clean signal. Our results show Hifi-CoGAN was able outperform Hifi-GAN many narrowband (signals with limited range frequencies) terms Short-Term Objective Intelligibility (STOI) Perceptual Evaluation Speech Quality (PESQ) metrics. did not perform as well wideband wider such white noise, so future work must be done these signals.