Suppressing Residual Acoustic Echo Based on Deep BiGRU with Attention
Chenggang Zhang, Xiaoqiang Wu · 2020
A residual echo suppressor (RES) aims to suppress the residual echo and nonlinear components in the output of a linear acoustic echo canceler (AEC). Traditional adaptive RES typically estimates the residual acoustic echo path from the inputs that are both the far-end speech and the error signal computed by the AEC, and derives the RES filter coefficients accordingly. However, the near-end speech is severely distorted when the residual echo attenuated in double-talk situations. In this paper, we propose a new supervised learning RES based on deep Bidirectional gated recurrent unity (BiGRU) with attention mechanism (BiGRU-Att). First, the BiGRU can effectively extract the inherent temporal feature of input signals, then attention mechanism assigns the corresponding weight to the extracted feature, taking an ideal ratio mask (IRM) as the learning target of the model. The experimental results show that compared with the deep neural network (DNN) -based RES, the proposed BiGRU-Att can not only suppress residual echo effectively, but also clearly restore near-end speech.