Research on Speech Enhancement Algorithm Based on SA-Unet

Most of the traditional speech enhancement algorithms are studied in the speech and audio domain. However, the suppression of noise by these methods is not obvious, and the effect of extracting pure speech is not good. In recent years, deep learning has been effectively developed based on its strong learning ability and is being used by more and more researchers. In view of this, this paper proposes a deep learning network model for research in the time domain, which is a SA-Unet network model that combines self-attention and Unet network structure and uses it for speech separation tasks. The model adds a self-attention mechanism between up-sampling and down-sampling in the Unet structure to enhance the perception between contexts, thereby more accurately separating the noise portion of the speech. Finally, the speech quality assessment scores are objectively obtained through the speech quality evaluation indicators and compared with the traditional speech enhancement algorithms. Experiments prove that SA-Unet has made outstanding contributions to speech enhancement technology.