Fusion Cross-Attention Convolutional Recurrent Network for Residual Echoes and Noise Suppression in Acoustic Echo Cancellation
Lei Zhang, Yi Zhou, Yu Zhao, Hongqing Liu · 2023
With the development of smart devices, voice interaction has become more frequent, and echoes and background noise are unavoidable during hands-free calls. Many previously proposed deep learning methods as a post-filtering module simply concatenate on error signals and estimated echoes as input features without considering the information interaction between them. In this paper, we propose the Cross Attention (CA) module to interactively fuse the feature information between the inputs and apply it to the Convolutional Recurrent Network structure (CRN). In addition, the fusion skip connection (FSC) module is also proposed to enable the decoder to simultaneously sense the fused information of the error signal and the estimated echo, which is benefit to improve the echo cancellation effect. Ablation experiments of the FCACRN model proposed in this paper on the synthetic test set show the effectiveness of the proposed module, which can eliminate echoes and noise more effectively. Meanwhile, the blind test set in the AEC Challenge also obtained excellent scores. The AECMOS of this method is 4.41 points, 0.54 points higher than the baseline, which proves the excellent performance of the proposed model.