Evaluation Platform and Benchmarks for Music Source Separation
Abstract
Sound demixing is the task of separating a music recording into its underlying stems, such as vocals, drums, bass, and other accompanying instruments. It has applications in music production, karaoke, remixing, audio restoration, and speech enhancement, among others. In this paper, we introduce MVSep Quality Checker, a publicly accessible evaluation platform for music source‑separation tasks. The platform provides open leaderboards, public submissions, standardized metric implementations, and a collection of existing and newly introduced benchmarks. We evaluate popular sound demixing models on these benchmarks and report results across multiple metrics. In addition to widely used metrics, we introduce Bleedless and Fullness, two new metrics that separately quantify unwanted bleed from non‑target sources and missing target‑source content.
© 2026 Roman Solovyev, Tatiana Habruseva, Jeremy Bonnot, Alexander Stempkovskiy, Ferdinand van Wijnen, published by Ubiquity Press
This work is licensed under the Creative Commons Attribution 4.0 License.