Speech Dereverberation Using Deep Learning Algorithm

Abstract

This paper focuses on speech derverberation using a single microphone. We investigate the applicability of fully convolutional networks (FCN) to enhance the speech signal represented by short-time Fourier transform (STFT) images in light of their recent success in many image processing applications. We present two variants: a "U-Net," which is an encoder-decoder network with skip connections, and a generative adversarial network (GAN) with the U-Net as the generator, which produces a more intuitive cost function for training. To assess our method, we used data from the REVERB challenge and compared our results to those of other methods tested under the same conditions. In most cases, we discovered that our method outperforms the competing methods.