This study proposes an improved UNet stereo echo cancellation method based on SCSconv and channel attention mechanism to address the problems of strong acoustic coupling, nonlinear echo path, large computational complexity, slow convergence, and insufficient speech fidelity of traditional adaptive filtering algorithms in car multi-channel audio systems. Firstly, clarify the principles of single channel and multi-channel echo cancellation, and analyze the mechanisms of echo and acoustic echo generation in vehicle transmission lines; Secondly, a lightweight deep learning model is constructed that integrates SCSconv spatial channel decoupling convolution and ECA/SKNet hybrid attention mechanism. Attention modules are embedded in the encoder and decoder layers to optimize speech feature extraction, reduce computational overhead, and suppress overfitting; Finally, comparative experiments were conducted under different reverberation times (RT60=0.3~0.9 s) and signal-to-noise ratios (-5 dB, 0 dB). The results showed that the proposed method achieved the highest echo return loss enhancement (ERLE) of 54.5 dB and the highest speech quality assessment (PESQ) of 2.80 in various vehicle acoustic environments. Its comprehensive performance was superior to traditional adaptive filtering and benchmark deep learning models, and it met the requirements of low latency (<20 ms) for real-time communication. It can effectively suppress vehicle reverberation, road noise, and multipath interference, providing an effective technical solution for quality control and stereo echo cancellation of intelligent vehicle audio systems.
