A music source separation method integrating time-frequency decoupling and mamba-based state space modeling

Area of Science:

• Intelligent Audio Processing
• Machine Learning for Music Information Retrieval

Background:

• Existing music source separation models face limitations in structural adaptability and balancing long-range dependencies with inference efficiency.
• Heterogeneous sound sources and entangled time-frequency representations pose significant challenges for current separation techniques.

Purpose of the Study:

• To propose a novel dual-path state space modeling architecture, MSNet, for advanced music source separation.
• To address the structural limitations and efficiency challenges in current music separation models.

Main Methods:

• Developed a dual-path state space architecture (MSNet) with decoupled temporal and frequency pathways.
• Incorporated Mamba-based state space units for multidimensional structural parsing of audio signals.
• Enhanced selective control and structural representation in time-frequency modeling.

Main Results:

• MSNet achieved state-of-the-art performance on the MUSDB18 dataset across multiple metrics.
• Demonstrated superior robustness and stability for complex sources like vocals and drums.
• Achieved a real-time factor (RTF) below 0.1, indicating practical applicability.

Conclusions:

• State space models are feasible for complex audio modeling tasks like music source separation.
• MSNet introduces a new architectural paradigm balancing accuracy and efficiency in music source separation.
• The proposed model offers a practical solution for real-time music source separation applications.