arxiv:2610.17821

FLOMSS: Generative Music Source Separation by Flow Matching

Published on Oct 14

· Submitted by

Shuyang Cui on Oct 26

Upvote

Authors:

Shuyang Cui ,

Abstract

Flow matching is investigated for four-source full-band stereo music separation, evaluating prediction targets, flow paths, modeling strategies, and consistency constraints with both signal-level and perceptual metrics.

Generated by thinkingmachines/Inkling-Small

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 distributions, 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 objective and subjective tests, highlighting the potential of flow matching for generative MSS. Demo: https://ik4sumii.github.io/FLOMSS/

View arXiv page View PDF Project page Add to collection

Community

ik4sumii

Paper author Paper submitter 1 day ago

FLOMSS systematically investigates flow matching for full-band stereo music source separation (MSS).

Systematic design study: We investigate prediction targets, flow paths, single- vs. multi-source modeling, and mixture consistency. Optimal Transport-CFM with x-prediction provides the most balanced performance across the evaluated metrics.
Beyond SDR: We evaluate MSS with signal-level, spectral, embedding-based, distributional, and subjective metrics, revealing a distributional–reconstruction trade-off: flow-matching models improve spectral and embedding-based similarity while generally sacrificing some SDR.
Strong generative MSS performance: FLOMSS-S outperforms prior generative baselines on most spectral and perceptual metrics.
Competitive subjective quality: Scaling to FLOMSS-M improves every reported objective metric over FLOMSS-S and ranks #2 overall in the listening test, despite its lower SDR than strong discriminative separators.
These results highlight flow matching as a promising direction for generative music source separation with competitive perceptual quality.