SPDMX Drops 254K Songs to Fix Music AI's Data Crisis

Music AI lacks the redistributable training data that accelerated progress in computer vision. Licenses usually prevent researchers from republishing commercial recordings, while accessible music datasets rarely include stems, the isolated instrument tracks required for source separation and other supervised tasks. The SPDMX paper addresses both constraints by rendering public-domain and CC0 scores into openly licensed audio, with aligned MIDI and metadata.

The release contains 253,982 songs, 490,159 mono stems, and 5,469 song-hours of audio sampled at 44.1 kHz. Its multitrack subset covers 88,403 songs and 2,734 song-hours, giving researchers two or more aligned stems for roughly one-third of the corpus. The authors describe it as the largest openly licensed synthetic multitrack collection by song-hours available at release.

Redistributable stems are the bottleneck

Source-separation systems learn to recover individual parts from a complete mix, so training requires both the mixture and its isolated targets. MUSDB18, the standard benchmark, provides 150 professionally produced songs and roughly 10 hours of music. Its realistic recordings support evaluation, but its size constrains data-hungry models.

Slakh2100 expands the synthetic option to 2,100 songs and about 145 hours, although reproducing its rendering stack requires proprietary virtual instruments. Pipelines built entirely from free SoundFonts are easier to distribute, but simple one-pass rendering often produces an obvious synthetic timbre that can weaken performance on real recordings.