Music denoising aims to filter out unwanted noise while preserving the original quality of the music signal. In particular, noise generated in crowded environments such as pubs or streets varies widely in both type and intensity. In these settings, conventional denoising techniques often fail to perform effectively, leading to a degradation in music quality. There are a few related works, such as Band-Split RNN (BSRNN), typically involve dividing the signal into frequency bands and processing each band individually. However, they face several limitations. For instance, BSRNN’s high model complexity results in increased computational costs, and its reliance on manual frequency band selection can lead to suboptimal performance. Furthermore, it struggles with separating overlapping frequency bands in complex environments. To overcome these limitations, this paper proposes the MusicBsECA method, which combines Efficient Channel Attention (ECA) with a band-splitting approach. Our method reduces model complexity and computational costs by focusing on critical features through ECA, while also leveraging LSTM and Transformer networks to capture both local and global dependencies in the signal. This custom approach allows for more precise separation of similar frequency bands and better performance in noise-heavy environments. The proposed method demonstrates superior performance in complex and noisy environments compared to existing techniques.