Application of recurrent U-net architecture to speech enhancement

Abstract

In this paper a recurrent U-net neural architecture is proposed to speech enhancement. The mentioned neural network architecture is trained to provide a mapping between a spectrogram of a noisy speech and both spectrograms of isolated speech and noise. Some key design choices are being evaluated in experiments and discussed, including: number of levels of the U-net, presence/absence of recurrent layers, presence/absence of max pooling layers as well and upsampling algorithm used in decoder part of the network.