Open Access Strategies for Music Information Retrieval

This direction is split into other directions: audio data accessibility. Its last version stays readable and citable here.

Version 1 of 131 papers34 conceptsfrozen 2026-10-04

In most fields of machine learning, progress is fueled by massive, shared datasets like ImageNet or Common Crawl. In Music Information Retrieval (MIR), however, researchers hit a legal wall: copyright. Because commercial music is strictly protected, researchers cannot simply upload a terabyte of MP3s for the community to download. This "copyright barrier" has historically forced the field into a "reproducibility crisis," where models are trained on private collections that no one else can access.

To move forward, the community has developed a set of clever workarounds—ranging from "open-source" music to sharing mathematical abstractions of sound rather than the sound itself.

The Reproducibility Crisis in Audio

When researchers cannot share their raw data, science suffers. If Fabian-Robert Stöter and colleagues develop a new source separation model like Open-Unmix, they need standardized "ground truth" audio to prove it works. Without shared datasets, it is impossible to know if a new algorithm is actually better or if it just happened to perform well on a specific, private library of songs.

Rachel M Bittner addressed this fragmentation by creating mirdata, a library that provides common loaders and validators for datasets. While it doesn't host the copyrighted audio, it ensures that if two researchers both own the same CD, they are at least processing the exact same bitstream in the same way.

Strategy 1: The Creative Commons Pivot

The most direct way to bypass copyright is to stop using commercial music. Researchers like Dan Stowell with freefield1010 and António Ramires with the Freesound Loop Dataset leverage Creative Commons (CC) audio.

By using platforms like Jamendo or the Free Music Archive (FMA), projects such as the MediaEval 2019 challenge can provide full, high-quality audio files to any participant. This allows for "end-to-end" learning where models look at raw waveforms, a luxury often denied to those working with Top 40 hits.

Strategy 2: Feature-Only Datasets

If you can't share the song, share its "fingerprint." Many modern datasets provide pre-computed features—mathematical summaries of the audio that are useful for machines but cannot be easily turned back into a listenable song for humans.

For example, Andres Ferraro and the team behind the Melon Playlist Dataset provided mel-spectrograms for 650,000 tracks. This allows researchers to train massive recommendation and tagging models on real-world Korean pop music without ever distributing a single copyrighted audio file. Similarly, AcousticBrainz, led by Alastair Porter and Dmitry Bogdanov, crowdsources these features from users' local libraries to build a global, open-access metadata repository.

Strategy 3: Remote Extraction and Expert Labeling

Sometimes, the value isn't in the audio itself, but in the "ground truth" labels provided by experts. Carlos N. Silla and Celso A. A. Kaestner demonstrated this with the Latin Music Database. Instead of relying on vague web-scraped tags, they used professional dance experts to label genres.

To share this knowledge legally, researchers often propose "remote metadata extraction" frameworks. In this setup, the algorithm travels to the data (on a secure server) rather than the data traveling to the researcher. This preserves the "tacit knowledge" of the experts while respecting the legal boundaries of the recordings.

The Symbolic Alternative

Finally, some researchers bypass the audio problem entirely by focusing on symbolic music—like MIDI or sheet music. Hao‐Wen Dong and Julian McAuley recently introduced MetaScore, a dataset of nearly a million scores. While symbolic data still faces licensing hurdles, it is often easier to distribute and annotate with AI-generated captions, providing a bridge for generative models to learn the "language" of music without the heavy legal and computational lift of raw audio.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_open_access_strategies_for_music_informa_0a28ee4d,
  title = {Open Access Strategies for Music Information Retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/open-access-strategies-for-music-information-retrieval/rdr_0a28ee4d6db34d54b8a260ee08c43e31},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_open_access_strategies_for_music_informa_0a28ee4d_v1,
  title = {Open Access Strategies for Music Information Retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/open-access-strategies-for-music-information-retrieval/rdr_0a28ee4d6db34d54b8a260ee08c43e31/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}