Speech Enhancement Using U-Net with Compressed Sensing

With the development of deep learning, speech enhancement based on neural networks had made a great breakthrough. The methods U-Net structure achieved good denoising performance. However, part them rely ordinary convolution operation, which may ignore contextual information and detailed features input speech. To solve this issue, many studies have improved model performance by adding additional network modules, such as attention mechanism, long short-term memory (LSTM), etc. In work, therefore, time-domain combines lightweight Shuffle Attention mechanism compressed sensing loss (CS loss) is proposed. dilated residual blocks are constructed used for down-sampling up-sampling in model. added to final output encoder focusing suppressing irrelevant audio information. A new defined using measurements clean enhanced sensing, it can further remove noise noisy experimental part, influence different functions proved through ablation experiments, effectiveness CS verified. Compared with reference models, proposed obtain higher quality intelligibility scores fewer parameters. When dealing outside dataset, still achieves performance, proves that not only achieve effect, but also has generalization ability.