Improve speech enhancement with Wave-USE-Net
Sinuo Qiao, Zhongdong Wu · 2023
At present, most deep learning speech enhancement models rely on spectrograms, and phase estimation is needed to restore the speech signal from spectrograms. In this paper, a speech enhancement model called Wave-USE-Net is proposed, which combines Squeeze-and-Excitation Networks(SENet) and Wave-U-Net. On the basis of Wave-U-Net directly estimating the waveform of the enhanced speech signal end-to-end, the SE block is used to suppress redundant features and effectively recover the enhanced speech signal. The experimental results show that compared with the baseline Wave-U-Net model, the proposed Wave-USE-Net model improves PESQ, CSIG, CBAK, COVL and SSNR by 7.5%,7.1%,1.2%,8.5% and 1.1% respectively on the VCTK dataset. The enhancement effect of the proposed model is better than that of other speech enhancement models based on Wave-U-Net, and the enhanced speech signal is closer to clean speech.