Million Song Dataset
LabROSA (Columbia University) and The Echo Nest
Audio features for a million songs, with no audio, frozen in December 2010.
Metadata and Echo Nest audio analysis (tempo, key, loudness, timbre and pitch segments) for one million popular tracks, plus community add-ons for lyrics, cover songs, tags, genres, and listener play counts (the Taste Profile subset). It was the music-information-retrieval field's first commercial-scale data set and is still used for recommendation and music-trend studies.
Access
- Cost
- FreeFree; full set is distributed via cloud snapshots, subset by direct download.
- License
- Provider's own termsEcho Nest API terms of use; MusicBrainz tags CC BY-NC-SA 2.0; add-on sets have their own terms
- Sign-up
- No sign-up
- How you get it
- Bulk downloadHDF5, SQLite, TXT
What's in it
- One row is
- One row per track
- Covers
- Mostly Western popular music
- Years
- Tracks released up to 2010; data collected December 2010
- Updated
- One-time release
- Size
- 1 million tracks; about 280 GB for the full set, with a 10,000-track subset
Quality
Foundational and well documented, but frozen in 2010, without audio, and built on an API that no longer exists.
- No audio, only derived features
- Nothing after 2010
- Some track-to-metadata matches are known to be wrong
Connects to
- MusicBrainz Database: Tracks carry MusicBrainz identifiers for richer release and artist metadata
- AudioSet: AudioSet's music-genre clips offer raw-audio counterparts to MSD's features
Questions it could answer
- Did popular music get louder between 1960 and 2010?
- Which audio features best predict a song's play count in the Taste Profile data?