Full-band stereo source separation

FLOMSS: Generative Music Source Separation by Flow Matching

Code repository ↗
Read abstract

Music source separation (MSS) has been conventionally addressed by discriminative methods. Recently, flow matching has emerged as a strong generative paradigm for modeling complex data distri- butions, with successful applications across diverse domains. These advances motivate us to systematically investigate flow matching for MSS in a four-source (Vocals, Bass, Drums, Other) setting in full-band stereo. Specifically, we investigate prediction targets, flow paths, single-source versus multi-source modeling, and mixture consistency constraints. Beyond the signal-level metrics commonly used for MSS, such as signal-to-distortion ratio, we further evaluate MSS methods using embedding-based metrics, as well as a listening test, providing a comprehensive benchmark. On MUSDB18-HQ, our proposed model exhibits competitive performance in both objec- tive and subjective tests, highlighting the potential of flow matching for generative MSS.

Evaluation results

Comprehensive evaluation

Subjective Quality

Mean subjective score with 95% confidence interval

Best Second best Higher is better ↑
Mean subjective scores and 95% confidence intervals by source
Method Params (M) Average ↑ Rank ↓ Vocals ↑ Bass ↑ Drums ↑ Other ↑

Mean subjective scores are reported across individual sources. Overall ranks exclude reference conditions and systems without results for all four sources.

Objective Metrics

Best Second best

Average across sources

Per-source results

↑ higher is better; ↓ lower is better. SDR, SIR, and SAR are reported in dB. FD is evaluated independently in each embedding space.

Interactive listening comparison

Choose a source, then switch systems at the same timestamp for a time-aligned A/B workflow.

00:00 / 00:00