An End-to-End Speech Enhancement Method Combining Attention Mechanism to Improve GAN

Abstract

Current Generative Adversarial Networks only rely on convolution operations when dealing with speech tasks, ignoring the dependencies between time series and have limited learning ability so that there is still obvious residual noise in the enhanced speech. To solve this problem, an end-to-end speech enhancement method combining attention mechanisms to improve GAN is proposed to apply a combined attention mechanism fusing channel and space between convolutional layers of SEGAN to obtain more contextual information of speech in both channel and space dimensions and extract more accurate feature information. Experimental results demonstrate that the method outperforms the baseline model in both speech quality and intelligibility. The experimental data show that under different signal-to-noise ratios, the perceptual speech quality assessment (PESQ) is improved by an average of 25.72%, and the objective short-term object intelligibility (STOI) is improved by an average of 1.68%.