Residual Echo Suppression using Spatial Feature for Stereo Acoustic Echo Cancellation

There have been some advances in deep learning based stereo-AEC (SAEC) systems in recent years. However, most of the studies focused on solving the non-uniqueness problem and did not explore the impact of spatial cues on SAEC. In this paper, we propose a composite SAEC system that combines conventional adaptive filters and Wiener filters with a NN-based residual echo suppression (RES) module. We adopt generalized cross correlation (GCC) as an additional input to allow the RES module to better analyze the embedded spatial cues. Experimental results show that the addition of GCC stably improves system performance with a little computational overhead. In addition, the proposed system achieves better results with much less computation than compared NN-based SAEC systems in all test conditions.