Speech Enhancement of Complex Convolutional Recurrent Network with Attention

Speech enhancement aims to separate pure speech from noisy speech, to improve speech quality and intelligibility. A complex convolutional recurrent network with a parameter-free attention module is proposed to improve the effect of speech enhancement. First, the feature information is enhanced by improving the convolutional layer of the encoding layer and the decoding layer. Then, the redundant information is suppressed by adding a parameter-free attention module to extract features that are more effective for the speech enhancement task, and the middle layer is selected for the bidirectional gated recurrent unit. Compared with the best of several baseline models, in the Voice Bank + DEMAND dataset, Perceptual Evaluation of Speech Quality (PESQ) increased by 0.17 (6.23%), MOS predictor of intrusiveness of background noise (CBAK) increased by 0.14 (4.34%), (MOS predictor of overall processed speech quality) COVL increased by 0.40 (12.42%), and (MOS predictor of speech distortion) CSIG index increased by 0.57 (15.28%). Experimental results show that the proposed approach has higher theoretical significance and practical value for actual speech enhancement.