FLOMSS: Generative Music Source Separation by Flow Matching

FLOMSS. Generative Music Source Separation by Flow Matching. By Shuyang Cui and twelve others, from Sony Group Corporation. Published on the fourteenth of October. No code repository is linked to it.

Music source separation, taking a finished mix and pulling the individual parts back out of it, has conventionally been treated as a discriminative problem. Meanwhile flow matching has emerged as a strong generative paradigm for modelling complex data distributions, with successful applications in many domains. Whether that generative approach suits separation, and which of its many design choices matter for it, had not been worked through systematically.

This work does that investigation, in a demanding setting: full-band stereo, with four sources to recover, namely Vocals, Bass, Drums, and Other. FLOMSS is the generative separation model that comes out of it.

Four design questions are examined. What the model should be trained to predict. Which flow path to use. Whether to model each source on its own or all sources together. And whether to impose a mixture consistency constraint. Evaluation is broadened too. Beyond the signal-level metrics usually used for separation, such as signal-to-distortion ratio, the authors add embedding-based metrics and a listening test, and present the combination as a comprehensive benchmark.

On MUSDB18-HQ, the proposed model performs competitively in both the objective measures and the subjective test. The abstract presents that as competitive rather than as a new best, and it reads the result as a showing of what flow matching could do for generative separation.

Adding a listening test changes who gets a say in what counts as a good separation. Signal-level scores, embedding scores and human listeners are three different judges, and this evaluation consults all three for a generative model. Which of them should lead whenever they disagree is a question the benchmark now makes it possible to ask.