Nonlinear Residual Echo Suppression using a Recurrent Neural Network

The acoustic front-end of hands-free communication devices introduces a variety of distortions to the linear echo path between the loudspeaker and the microphone. While the amplifiers may introduce a memory-less non-linearity, mechanical vibrations transmitted from the loudspeaker to the microphone via the housing of the device introduce non-linarities with memory, which are much harder to compensate. These distortions significantly limit the performance of linear Acoustic Echo Cancellation (AEC) algorithms. While there already exists a wide range of Residual Echo Suppressor (RES) techniques for individual use cases, our contribution specifically aims at a low-resource implementation that is also real-time capable. The proposed approach is based on a small Recurrent Neural Network (RNN) which adds memory to the residual echo suppressor, enabling it to compensate both types of non-linear distortions. We evaluate the performance of our system in terms of Echo Return Loss Enhancement (ERLE), Signal to Distortion Ratio (SDR) and Word Error Rate (WER), obtained during realistic double-talk situations. Further, we compare the postfilter against a state-of-the art implementation. Finally, we analyze the numerical complexity of the overall system.