Summary of The Inaugural Music Source Restoration Challenge

TL;DR

This work targets restoring original instrument stems from degraded mixes, addressing production effects and real-world degradations beyond standard music source separation. It introduces the inaugural MSR Challenge and the MSRBench dataset, with objective metrics including Multi-Mel-SNR, Zimtohrli, and FAD-CLAP and subjective MOS ratings on real degradations to evaluate both fidelity and perceptual quality. Five teams submitted results; the top xlancelab achieved MMSNR of $4.46$ dB and MOS-Overall of $3.47$, corresponding to relative improvements of $91\%$ and $18\%$ over the second place, respectively, with per-stem trends showing bass easier than percussion. The findings suggest that sequential/ensemble architectures with simple reconstruction losses outperform adversarial training, while data quality matters; the MSRBench dataset and baselines are publicly available to advance future MSR research.

Abstract

Music Source Restoration (MSR) aims to recover original, unprocessed instrument stems from professionally mixed and degraded audio, requiring the reversal of both production effects and real-world degradations. We present the inaugural MSR Challenge, which features objective evaluation on studio-produced mixtures using Multi-Mel-SNR, Zimtohrli, and FAD-CLAP, alongside subjective evaluation on real-world degraded recordings. Five teams participated in the challenge. The winning system achieved 4.46 dB Multi-Mel-SNR and 3.47 MOS-Overall, corresponding to relative improvements of 91% and 18% over the second-place system, respectively. Per-stem analysis reveals substantial variation in restoration difficulty across instruments, with bass averaging 4.59 dB across all teams, while percussion averages only 0.29 dB. The dataset, evaluation protocols, and baselines are available at https://msrchallenge.com/.