Speech noise reduction algorithm based on CA-DCDCCRN

To extract more useful feature information from the noisy speech spectrogram, this paper proposes a Deep Complex Densely Connected Convolutional Recurrent Network based on Coordinated Attention Mechanism (CA-DCDCCRN) speech reduction algorithm. The CA-DCDCCRN model incorporates Dense Convolution (DC) and Complex Dense Convolution (CDC) on the basis of Deep Complex Convolution Recurrent Network (DCCRN) to improve the modeling ability of models using contextual information, and adding the Coordination Attention (CA) module to the encoder to improve the attention to noisy spectrogram features. The experimental results show that the algorithm ensures the deep supervision and feature reuse capability of the network and reduces the loss of speech detail information; compared with the speech noise reduction methods of SEGAN, WaveU-net and DCCRN, the average improvement of PESQ is 33.33%, 30.8% and 9.9%, and the average improvement of STOI is 20.63%, 19.10% and 4.35%, which provides a more effective method for speech processing in complex environments.