Attention Wave-U-Net for acoustic echo cancellation

Abstract

In this paper, a Wave-U-Net based acoustic echo cancellation (AEC) with an attention mechanism is proposed to jointly suppress acoustic echo and background noise. The proposed approach consists of the Wave-U-Net, an auxiliary encoder, and an attention network. In the proposed approach, the Wave-U-Net yields the estimated near-end speech from the mixture, the auxiliary encoder extracts the latent features of the far-end speech, among which the relevant features are provided to the Wave-U-Net by using the attention mechanism. With the attention network, the echo can be effectively suppressed from the mixture. Experimental results on TIMIT dataset show that the proposed approach outperforms the existing methods in terms of the echo return loss enhancement (ERLE) for the single-talk period and the perceptual evaluation of speech quality (PESQ) score for the double-talk period. Furthermore, the robustness of the proposed approach against unseen noise condition is also validated from the experimental results.