Datasets:

MSRChallenge
/

2025TestSet

pretty_name: MSR Challenge 2025 Evaluation Set
license: cc-by-nc-4.0
task_categories:
  - audio-to-audio
language:
  - zxx
size_categories:
  - 1K<n<10K
tags:
  - music
  - audio-restoration
  - source-separation

MSR Challenge 2025 Evaluation Set

This repository contains the held-out evaluation audio distributed for the inaugural Music Source Restoration (MSR) Challenge associated with ICASSP 2026. The task is to recover an unprocessed musical source from a professionally produced or otherwise degraded mixture.

The release has 1,496 stereo mixture clips across eight instrument classes. Reference targets are available for 1,128 examples; the remaining 368 examples are mixture-only because the organizer archive did not include references for the cylinder, live, or radio subsets.

Dataset structure

./
├── metadata.jsonl
├── mixtures/
│   └── {instrument}/{file_id}.flac
└── targets/
    └── {instrument}/{file_id}.flac

All audio is stereo FLAC at 48 kHz and 10 seconds long. Mixtures and targets retain their original encoded bit depths.

Subset	Mixtures	Targets	Description
`non-blind`	1,000	1,000	Organizer non-blind evaluation material
`streaming`	128	128	Lossy-codec conditions
`cylinder`	112	0	Historical cylinder recordings
`live`	128	0	Live/acoustic degradation condition
`radio`	128	0	Radio degradation condition
Total	1,496	1,128

Instrument counts are 189 each for Bass, Drums, Guitars, Keyboards, Orchestral Elements, Percussions, and Vocals, and 173 for Synthesizers.

Metadata fields

file_id: stable identifier used by the organizer archives.
mixture_file_name: relative path to the input audio.
target_file_name: relative path to the reference audio, or null when unavailable.
has_target: whether a reference is included.
subset: non-blind, streaming, cylinder, live, or radio.
instrument: target instrument class.
augmentation_type, augmentation_code: codec/augmentation description when present.

Intended use

This dataset is intended for evaluation of music source restoration and related source-separation or audio-restoration systems. Do not treat the mixture-only examples as having negative or silent targets. Users should report results separately by subset and instrument where possible.

Citation

Please cite the challenge summary when using this evaluation set. The original MSR task paper and the related MSRBench paper are also included below.

@inproceedings{zang2026msrchallenge,
  title     = {Summary of the Inaugural Music Source Restoration Challenge},
  author    = {Zang, Yongyi and Hai, Jiarui and Ge, Wanying and Kong, Qiuqiang and Dai, Zheqi and Wang, Helin and Mitsufuji, Yuki and Plumbley, Mark D.},
  booktitle = {ICASSP 2026--2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages     = {21853--21855},
  year      = {2026},
  doi       = {10.1109/ICASSP55912.2026.11462762}
}

@inproceedings{zang2025music,
  title     = {Music Source Restoration},
  author    = {Zang, Yongyi and Dai, Zheqi and Plumbley, Mark D. and Kong, Qiuqiang},
  booktitle = {2025 IEEE International Workshop on Multimedia Signal Processing (MMSP)},
  pages     = {138--143},
  year      = {2025},
  doi       = {10.1109/MMSP64401.2025.11324269}
}

@inproceedings{zang2026msrbench,
  title     = {MSRBench: A Benchmarking Dataset for Music Source Restoration},
  author    = {Zang, Yongyi and Hai, Jiarui and Ge, Wanying and Dai, Zheqi and Wang, Helin and Mitsufuji, Yuki and Kong, Qiuqiang and Plumbley, Mark D.},
  booktitle = {Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year      = {2026}
}

License

This dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License.