Nonlinear Residual Echo Suppression using a Recurrent Neural Network

The acoustic front-end of hands-free communication de-vices introduces a variety of distortions to the linear echo pathbetween the loudspeaker and the microphone. While the ampli-fiers may introduce a memory-less non-linearity, mechanical vi-brations transmitted from the loudspeaker to the microphone viathe housing of the device introduce non-linarities with memory,which are much harder to compensate. These distortions signif-icantly limit the performance of linear Acoustic Echo Cancella-tion (AEC) algorithms. While there already exists a wide rangeof Residual Echo Suppressor (RES) techniques for individualuse cases, our contribution specifically aims at a low-resourceimplementation that is also real-time capable. The proposedapproach is based on a small Recurrent Neural Network (RNN)which adds memory to the residual echo suppressor, enabling itto compensate both types of non-linear distortions. We evaluatethe performance of our system in terms of Echo Return Loss En-hancement (ERLE), Signal to Distortion Ratio (SDR) and WordError Rate (WER), obtained during realistic double-talk situa-tions. Further, we compare the postfilter against a state-of-theart implementation. Finally, we analyze the numerical complex-ity of the overall system.