SPDMX: Public-Domain Multitrack Stem Audio at Scale

Phillip Long, Eduardo Escoto, Julian McAuley, Zachary Novack

University of California, San Diego

About & download Listen Blog Code PDMX

Recorded multitrack stems remain scarce for modern separation and generative audio models, while large synthesized alternatives often depend on closed commercial instruments. SPDMX is a public-domain multitrack stem corpus from PDMX: scores are cleaned, rendered with open soundfonts and selective MIDI-DDSP, then loudness- and velocity-aware mixed into mono stems and mixes at 44.1 kHz—sharing song identity with PDMX so symbolic metadata stays usable alongside the audio. Song-hours count mix duration once per song.

chunk_N / song_id /
  • 0.flac … K.flac
  • mix.flac
  • mix.mid

44.1 kHz · mono stems · mono mix · FLAC · −23 LUFS

Citation & license

Cite SPDMX if you use the audio (stems or mixes). Also cite PDMX if you use anything beyond the paired MIDI shipped with SPDMX (for example MusicXML, PDMX metadata, or other PDMX-only fields).

SPDMX

@inproceedings{long2026spdmx,
  title={{SPDMX}: Public-Domain Multitrack Stem Audio at Scale},
  author={Long, Phillip and Escoto, Eduardo and McAuley, Julian and Novack, Zachary},
  booktitle={ICASSP},
  year={2026},
  note={to appear}
}

PDMX

@inproceedings{long2024pdmx,
  author={Long, Phillip and Novack, Zachary and Berg-Kirkpatrick, Taylor and McAuley, Julian},
  booktitle={ICASSP 2025},
  title={{PDMX}: A Large-Scale Public Domain MusicXML Dataset for Symbolic Music Processing},
  year={2025},
  pages={1-5},
  doi={10.1109/ICASSP49660.2025.10890217}
}

Release license: CC BY 4.0 (PDMX scores are public domain; see the dataset LICENSE). This site is the primary hub for demos, plots, and experiment write-ups.