DBAUNet: Dual-branch attention U-Net for time-domain speech enhancement
Bengbeng He, Kai Wang, Wei‐Ping Zhu · TENCON 2022 - 2022 IEEE Region 10 Conference (TENCON) · 2022
Recently, many speech enhancement methods in-volve attention mechanism to learn long-term dependencies of speech signals. And the U-Net structure is widely used for extracting hierarchical features. In this paper, we propose a dual-branch attention U-Net for speech enhancement in the time domain, named as DBAUNet, which consists of a convolutional en-coder, a dual-branch attention block and a convolutional de-coder. The encoder is used to extract the compressed features from input noisy speech. Then, the dual-branch attention block employs spatial-wise and channel-wise attention to extract the spatial and channel information of speech sequences in parallel, which are fused to learn the contextual information. Then a de-coder which has a symmetric structure with the encoder, is adopted to reconstruct the enhanced speech waveform. Experimental results on the benchmark dataset demonstrate that our proposed DBAUNet achieves a comparable performance to existing models while involving the fewest model parameters.