A Creative Commons archive offers cleared audio for MIR research

The audio task

A researcher or model builder who wants training or evaluation audio without the provenance questions attached to scraped web recordings needs a dataset where the licence on each track is documented, not assumed.

What the documents show

The FMA dataset repository documents a collection built from the Free Music Archive, totalling 106,574 tracks, about 917 gibibytes and 343 days of audio, spanning 161 genres contributed by 16,341 artists across 14,854 albums, and offered in several preset sizes from an 8,000-track “small” split to a “full,” untrimmed release. The repository states audio is distributed “under the license chosen by the artist,” predominantly Creative Commons licences, while the accompanying metadata is released under CC BY 4.0 and the processing code under the MIT licence, and that pre-computed audio features are included to support genre recognition, classification, and other music-information-retrieval tasks. The repository also flags known issues and corrections in its wiki rather than presenting the dataset as error-free. Separately, the live Free Music Archive site, as retrieved on 16 September 2026, describes itself as “powered by Tribe of Noise” and offers tracks framed as “royalty free” for use in videos and podcasts, wording that sits alongside, rather than restates, the dataset repository's more specific Creative Commons framing.

Rights status

The dataset repository is explicit that licensing is per-track and artist-chosen, not a single blanket grant covering the whole archive, and that copyright responsibility rests with the contributing artist rather than the dataset's maintainers. The live site's “royalty free” language, read on its own, is less specific than the dataset documentation's Creative Commons framing, and the two should not be treated as interchangeable when a specific track's reuse terms matter.

What to check before you use it

This is an editorial checklist. Before training on or redistributing FMA-derived audio, check the specific Creative Commons variant attached to each track, since Creative Commons covers licences ranging from attribution-only to non-commercial and no-derivatives, and the dataset's own per-track metadata is where that detail lives, not the summary statistics. Confirm whether a given project used the original research dataset's documented licences or sourced audio through the current freemusicarchive.org site under its “royalty free” framing, since the two are not stated in identical terms. Review the repository's listed known issues before relying on any single track's metadata as authoritative.

• Does the specific track's Creative Commons variant permit the intended use, including commercial and derivative uses?
• Was the audio sourced from the documented research dataset or from the current website's own licensing framing?
• Has the repository's errata wiki flagged an issue with the specific tracks or metadata being relied on?

FMA's value is not that every licence question disappears, but that the questions have documented, checkable answers at the track level, which is the property that makes it useful as a rights-cleared point of comparison against training data whose provenance is unknown.