基于CBAM-CRN的面向会议场景的多通道回声消除模型

Multi-Channel Echo Cancellation Model for Conference Scenarios Based on CBAM-CRN

本文研究了基于深度学习的多通道回声消除方法,提出了基于卷积块注意力模块(CBAM)融合卷积循环网络(CRN)的多通道回声消除方法。该方法利用U型网络的特征提取能力和LSTM网络处理时序信号的优势,结合了时频掩蔽算法和稀疏自适应归一化处理,同时融合了通道注意力和空间注意力联合机制,该混合域注意力能够有效地捕获关键特征并抑制无关特征。实验表明,CBAM-CRN方法在多种通话模式下均优于自适应滤波和其他深度学习方法,有效提高了远场免提通话的语音质量。

In this paper, we study the multi-channel echo cancellation method based on deep learning for acoustic echo problem, and propose a multi-channel echo cancellation method based on convolutional block attention module (CBAM) and convolutional recurrent network (CRN). This method takes advantage of the feature extraction ability of U-Net and the advantages of LSTM network in processing time series signals, combines the time-frequency masking algorithm and sparse adaptive normalization processing, and fuses the channel attention and spatial attention joint mechanism, the hybrid domain attention can effectively capture key features and suppress irrelevant features. Experimental results show that the CBAM-CRN method is superior to adaptive filtering and other deep learning methods in various call modes, and effectively improves the voice quality of far field hands-free calls.