Speech Enhancement Method Based on Generative Adversarial Network and Convolutional Block Attention Module

Current generative adversarial networks ignore the importance of information in space and among different convolutional channels when dealing with speech enhancement tasks, and do not adequately resolve features at different scales. Based on this, a speech enhancement method based on generative adversarial networks and convolutional attention modules is proposed. The generator can parse features at different time scales and a convolutional block attention module is introduced in the downsampling block of the generator. It consists of two sub-modules, the channel attention mechanism and the spatial attention mechanism, to extract more critical information in both channel and spatial dimensions. Using the method of training generative adversarial networks, the scores of the noise-canceled speech by this method are significantly improved in all the metrics when compared to the Wave-U-net model on the publicly available dataset. The improvement is 11.25%, 3.7%, and 8.8% on the three main metrics, namely PESQ, CBAK, and CSIG, respectively.