<p>The separation and identification of bird sounds are crucial for scientific research and ecological conservation. However, environmental noise and the distance between birds and recording devices pose significant challenges to sound separation. To address these issues, we propose BirdDenoiser, a time-domain bird sound denoising network based on optimized modern convolution. BirdDenoiser utilizes dilated convolution blocks to capture local audio features effectively, adapting to the diverse characteristics of bird sounds. By adjusting the position of depthwise convolutions, we reduce computational costs while maintaining high performance. Enlarging the kernel size further enhances the receptive field, improving the capture of local features from a multi-scale perspective. Additionally, we improve the activation function and incorporate global response normalization to further enhance the model’s performance. We constructed the BirdCLEFmix dataset and conducted extensive experiments to evaluate BirdDenoiser’s performance and efficiency. The experimental results demonstrate that BirdDenoiser significantly improves bird sound denoising while offering a competitive advantage in computational efficiency.</p>