Attention-Based Dual Stream Interactive Network for Nonlinear Residual Echo Suppression

Acoustic echo cancellation systems commonly employ linear adaptive filters to identify speaker-to-microphone echo paths. However, accurately estimating the echo path is challenging due to the nonlinear relationship between the far-end signal and the echo signal, leading to residual echo generation. Consequently, a post-suppression module is crucial for sufficient echo attenuation. Conventional deep learning-based methods for residual echo suppression (RES) rely on linear operations, such as addition and concatenation, disregarding the contextual information needed to effectively fuse error and auxiliary signal features. In this paper, we propose a novel end-to-end method for RES, which introduces an attentional fusion module that aggregates global and local contexts, as well as dynamically calculates the fusion weights for different signal features, enabling the neural network to efficiently leverage the correlation between these signals. Experimental results demonstrate the superiority of the proposed method.