Sdgan: Improve Speech Enhancement Quality by Information Filter

作     者：Xiaozhou Guo Yi Liu Wenyu Mao Jixing Li Wenchang Li Guoliang Gong Huaxiang Lu 

作者机构：Institute of Semiconductors Chinese Academy of Sciences Beijing 100083 China University of Chinese Academy of Sciences Beijing 100089 China CAS Center for Excellence in Brain Science and Intelligence Technology Chinese Academy of Sciences Beijing 200031 China Semiconductor Neural Network Intelligent Perception and Computing Technology Beijing Key Lab Beijing 100083 China 

出 版 物：《Journal of Physics: Conference Series》 

年 卷 期：2021年第1871卷第1期

学科分类：07[理学] 0702[理学-物理学] 

摘      要：The speech denoising model based on adversarial generative network has achieved better results than the traditional machine learning model. In this paper, for the short cut connection in the generator, we discuss its influence on the information transfer between encoder and decoder, and propose SDGAN at target. SDGAN sets linear and convolution filters in the short cut connection which adaptively learn the optimal information processing. The information filter still enables the generator to solve the gradient vanishing problem, and it can also avoid information redundancy and improve expression ability. In addition, SDGAN replaces the L1 regularization term in loss function with the L2 regularization term, which not only makes the output speech of the generator closer to the clean speech, but also avoids sparsity. In the experiments, SDGAN significantly performs better than other traditional GAN in five performance metrics (such as PESQ), and the effect of convolution filter is better than that of linear filter.