Temporal Convolutional Network for Acoustic Echo Cancellation in Double-Talk Scenarios

Abstract

In communication systems, when the loudspeaker and the microphone are coupled together, it creates acoustic echoes. With the increasing demand for mobile communication and online conference, it is urgent to solve the problem of acoustic echo cancellation (AEC) in communication systems. Due to the existence of nonlinear distortion, background noise and other reasons, traditional AEC methods can no longer solve the problem of echo cancellation well. Although some traditional methods consider the problem of nonlinear distortion, the effect of echo suppression is still not ideal. In this paper, we propose an echo cancellation method based on frequency domain mask, which is defined as a supervised speech separation problem. The use of the temporal convolutional network and optimal ratio mask to obtain the predicted mask, as well as the use of SISNR as the loss function, have been shown to effectively reduce echo in double-talk, nonlinear distortion, and background noise. This method is a significant advancement in the field of AEC and can be used in for mobile communication and online conference.