FLOMSS applies flow matching to four-source music separation
Original title: FLOMSS: Generative Music Source Separation by Flow Matching
The model separates vocals, bass, drums, and other in full-band stereo and is competitive on MUSDB18-HQ in objective and listening tests.
FLOMSS studies generative music source separation with flow matching instead of the usual discriminative setup. The work covers a four-source full-band stereo task (vocals, bass, drums, other) and compares prediction targets, flow paths, single-source versus multi-source modeling, and mixture consistency constraints. Evaluation uses signal-to-distortion ratio, embedding-based metrics, and a listening test. On MUSDB18-HQ, the proposed model is described as competitive in both objective and subjective tests.
- Sources are vocals, bass, drums, and other in full-band stereo.
- Metrics go beyond signal-to-distortion ratio to embeddings and a listening test.
- A public demo is linked with the paper.
From the original article
1 min read at huggingface.co
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.
By Shuyang Cui, Zhi Zhong, Keisuke Toyama · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.