Speech Enhancement using Convolutional Neural Network with Skip Connections
Yupeng Shi, Weicong Rong, Nengheng Zheng · 2018
Eliminating the negative effect of adverse environmental noise has been an intriguing but challenging research topic for speech enhancement and speech recognition. Numerous techniques using neural networks have achieved favorable denoising performance in recent years. This paper presents a denoising architecture of Convolutional Neural Network (CNN) and its derivates with skip connection(s) for speech denoising. The skip connection forces the CNN to learn the residual error between the noisy speech and the clean speech and the clean speech is finally estimated by subtracting the learned residual error from the noisy input. Experimental results demonstrate that the proposed CNN structure provides better denoising ability than Wiener filtering in noise reduction even when the model was tested using the data and noise set not included in the training set. It is also shown that the proposed model with single skip connection achieves considerably better performance under several real noise conditions even at lower Signal-to-Noise Ratio (SNR).