Speech Dereverberation Using Fully Convolutional Networks

Speech derverberation using a single microphone is addressed in this paper. Motivated by the recent success of the fully convolutional networks (FCN) in many image processing applications, we investigate their applicability to enhance the speech signal represented by short-time Fourier transform (STFT) images. We present two variations: a “U-Net” which is an encoder-decoder network with skip connections and a generative adversarial network (GAN) with U-Net as generator, which yields a more intuitive cost function for training. To evaluate our method we used the data from the REVERB challenge, and compared our results to other methods under the same conditions. We have found that our method outperforms the competing methods in most cases.

I. Introduction

Reverberation, resulting in from multiple reflections from the rooms facets and objects, degrade the speech quality, and in severe cases, the speech intelligibility, especially for hearing impaired people. The success rate of automatic speech recognition (ASR) systems may also significantly deteriorate in reverberant conditions, especially in cases of mismatch between the training and test phases. Reverberation is the result of convolving an anechoic speech utterance by a long acoustic path. The output signal suffers from overlap-and self-masking effects that may deteriorate the speech quality [1]. These are often manifested as “blurring” effects on the short-time Fourier transform (STFT) images. A plethora of methods for speech dereverberation using both single-and multimicrophone exists [2].