Residual Unet with Attention Mechanism for Time-Frequency Domain Speech Enhancement

Eliminating the negative effects of background environmental noise is an interesting and challenging task in audio processing. In recent years, denoising technology based on neural networks (NN) has achieved good performance. In particular, the structure based on the convolutional encoder and decoder has been proven to achieve good enhancement effects. On this basis, this paper proposes a residual unet structure combined with the attention mechanism. Effectively reduce the impact of gradient disappearance on network training, and improve the semantic gap between encoder output and decoder output due to unet shortcut connections. The experimental results show that compared with the DNN baseline and unet network, the enhanced voice quality has been significantly improved.