Effects of Skip Connections in CNN-Based Architectures for Speech Enhancement

摘要

Eliminating the negative effect of adverse environmental noise has been an intriguing and challenging task for speech technology. Neural networks (NNs)-based denoising techniques have achieved favorable performance in recent years. In particular, adding skip connections to NNs has been demonstrated to significantly improve the performance of NNs-based speech enhancement systems. However, in most of the studies, the adding of skip connections was kind of tricks of the trade and lack of sufficient analyses, quantitatively and/or qualitatively, on the underlying principle. This paper presents a denoising architecture of Convolutional Neural Network (CNN) with skip connections for speech enhancement. Particularly, to investigate the inherent mechanism of NNs with skip connections in learning the noise properties, CNN with different skip connection schemes are constructed and a set of denoising experiments, in which statistically different noises being tested, are presented to evaluate the performance of the denoising architectures. Results show that CNNs with skip connections provide better denoising ability than the baseline, i.e., the basic CNN, for both stationary and nonstationary noises. In particular, benefit by adding more sophisticated skip connections is more significant for nonstationary noises than stationary noises, which implies that the complex properties of noise can be learned by CNN with more skip connections.