Research on an Improved TF-GridNet-Based Method for Music Source Separation

音乐源分离任务旨在从混合音频信号中提取出不同的音源成分(如人声、鼓、贝斯等)，在音乐制作、智能语音、音频编辑等领域具有广泛的应用价值。TF-GridNet是近年来提出的一种融合时间与频率建模的深度网络结构，具备良好的建模能力。文章在TF-GridNet的基础上进行结构性改进，提出三项关键优化策略：1) 在时间建模路径中引入轻量通道注意力机制，提升模型对关键信号的响应能力；2) 在 GridBlock模块中引入残差门控机制，增强特征流动与融合灵活性；3) 在解码器部分设计多尺度重建路径，以提升高频细节还原效果。实验结果表明，改进后的模型在多源类别上优于原始TF-GridNet，并具备更优的感知质量与计算效率。Music source separation aims to effectively extract different components (such as vocals, drums, bass, etc.) or vocal components from a mixed audio signal, which has significant application value in intelligent audio processing, music production, and speech recognition. TF-GridNet is a recently proposed network architecture that combines time and frequency modeling and has demonstrated strong separation capability. Based on TF-GridNet, this paper introduces three structural improvements: 1) Introduces a lightweight channel attention mechanism in the time modeling path to enhance the model’s ability to respond to critical signals; 2) Introduces a residual gating mechanism in the GridBlock module to enhance the flexibility of feature flow and fusion; 3) Designs multi-scale reconstruction paths in the decoder section to improve high-frequency detail restoration effects. Experiments conducted on the MUSDB18 dataset show that the improved model outperforms the original TF-GridNet on multiple source types, with better modeling efficiency and perceptual quality.