Deep Neural Network Based Regression Approach for Acoustic Echo Cancellation

An acoustic echo canceller (AEC) aims to remove the acoustic echo in the mixture signal received by the near-end microphone. The conventional method uses an adaptive finite impulse response (FIR) filter to identify a room impulse response (RIR)which is not robust to various wild scenarios. In this paper, we propose a deep neural network-based regression approach that directly estimates the amplitude spectrum of the near-end target signal from features extracted from the mixtures of near-end and far-end signals. Depend on the powerful modelling and generalizing ability of deep learning, the complex echo signal can be well eliminated. Experimental results show the effectiveness of the proposed method for echo removal in double-talk, background noise, RIR variation and nonlinear distortion scenarios. In addition, the proposed method generalizes well to real-life acoustic echoes recorded in vehicles.