End‐to‐End Acoustic Echo Cancellation Based on Time‐Domain Neural Networks

Recently deep learning has become an important choice for acoustic echo cancellation, which mostly performs in the short‐time Fourier transform domain, but it may be limited by frame size selection and lack of accurate phase information. To overcome these limitations, in this paper, two improved Wave‐U‐Net temporal network models are proposed for single‐channel acoustic echo cancellation. Our objective is to enhance the model's ability to capture speech signal features by introducing a high‐resolution feature extraction module after each codec layer, and to capture the echo feature information more accurately by establishing a dual‐path fusion echo attention mechanism. For scenarios with low signal‐to‐echo ratios, we propose a second enhancement comprising two modules, including echo separation and residual feature fusion. Experimental results demonstrate that the proposed model achieves superior performance. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.