结合残差与双注意力机制的U-Net语音增强方法  

U-Net speech enhancement method combining residual and dual attention mechanism

作　　者：许春冬[1] 王磊 胡菁兰 闵源 徐锦武 XU Chun-dong;WANG Lei;HU Jing-lan;MIN Yuan;XU Jin-wu(School of Information Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学信息工程学院,江西赣州341000

出　　处：《计算机工程与设计》2024年第11期3383-3389,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61671442、11864016、11704164);江西省科技厅重点研发计划一般基金项目(20202BBEL53006);江西省文化艺术科学规划项目一般基金项目(YG2017384);江西理工大学研究生创新基金项目(XY2022-S168)。

摘　　要：针对U-Net语音增强网络深层特征提取能力不足,以及编解码过程中特征信息丢失问题,提出一种结合残差与双注意力机制的DA-Res-Unet语音增强方法。将U-Net编解码部分设计为残差结构来深化网络,增强深层特征提取能力;在网络结构中构造双注意力机制,减少时频特征提取中的细节信息丢失;在网络中融入空洞空间金字塔池化结构,在低参数量情况下融合不同尺度上下文背景信息,提高模型特征捕获能力。实验结果表明,DA-Res-Unet网络模型在可见噪声测试集上的PESQ、STOI和LSD这3种评测指标取得了不同程度的提升,在未知噪声测试集上具备一定优势。Aiming at the insufficient deep feature extraction ability of U-Net speech enhancement network and the loss of feature information in the process of encoding and decoding,a DA-Res-Unet speech enhancement method combining residual and dual attention mechanism was proposed.The U-Net codec part was designed as a residual structure to deepen the network and enhance the deep feature extraction ability.A dual attention mechanism was constructed in the network structure to reduce the loss of detail information in time-frequency feature extraction.The atrous spatial pyramid pooling operation was applied in the network to fuse the background information of different scales under low parameter conditions to improve the feature capture ability of the model.Experimental results show that the proposed network model achieves different degrees of improvement in the three evaluation indexes of PESQ,STOI and LSD on the visible noise test set,and it has certain advantages in the unknown noise test set.

关 键 词：语音增强 深度学习 残差网络 特征提取 编解码结构 注意力机制 空洞空间池化金字塔 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TN912.35[自动化与计算机技术—控制科学与工程]