Loading datasets...
Release Date: 11/4/2026
Format: CSV, FLAC, MIDI
Size: 742.65 GB
SPDMX (Synthesized PDMX) renders the public-domain PDMX symbolic corpus into mono FLAC stems, optional linear-sum mixes, and dense corrected MIDI at 44.1 kHz (253,982 songs; ~5,469 mixture hours). Join to PDMX on song_id. Layout: chunk_N/<song_id>/{k.flac, mix.mid, mix.flac?}. Mozilla Data Collective hosts this corpus as two archives, parts 1 and 2. Download both and extract them into the same directory: - Part 1: https://mozilladatacollective.com/datasets/cmw25vhb8001r08mqv2rhzlvu - Part 2: https://mozilladatacollective.com/datasets/cmw25vlq9001c07nx0f2l090a Part 1 contains even chunk ids (chunk_0, chunk_2, ...) and part 2 contains odd chunk ids. The CSV indexes in each archive describe the full corpus. Then optionally run ./link_single_track_mixes.sh. Single-track songs (~65%) do not ship mix.flac (it would duplicate 0.flac). The mix CSV column still points at …/mix.flac. Resolve by (1) loading 0.flac when subset:multitrack is false / n_tracks == 1, or (2) running ./link_single_track_mixes.sh after extract to create mix.flac → 0.flac symlinks. Multitrack songs already include mix.flac. See README.md in the archive. Synthesis uses open FluidSynth soundfonts plus selective MIDI-DDSP / DDSP-Piano. License: CC BY 4.0. Demo: https://pnlong.github.io/SPDMX/
Licensing
Creative Commons Attribution 4.0 International (CC-BY-4.0)
https://spdx.org/licenses/CC-BY-4.0.htmlLicense and attribution. Use is subject to Creative Commons Attribution 4.0 International (CC BY 4.0).
Nature of the audio. The audio is machine-synthesized from symbolic scores. It is not a collection of human studio recordings. Do not present SPDMX stems or mixes as live or commercial studio performances of identifiable musicians.
Public-domain provenance. Underlying scores and public-domain designations originate in PDMX. We rely on PDMX’s public-domain labeling for those symbolic sources; SPDMX does not independently re-clear copyright for every MuseScore/PDMX item. Users remain responsible for complying with applicable law in their jurisdiction when redistributing or commercially exploiting derived works.
Misrepresentation.
Attribution. Do not remove or obscure required attribution to SPDMX / PDMX when sharing the data or models trained on it, beyond what CC BY 4.0 already requires.
Warranty of title. Do not assert that SPDMX’s authors warrant that every underlying score is public domain in every jurisdiction. Public-domain status is inherited from PDMX’s curation and labeling.
Ethical Review
No human-subjects protocol or IRB review was required for SPDMX audio creation: there were no participant recordings or personally identifying performance captures. Audio is synthesized from symbolic scores that PDMX curated and labeled as public domain. We do not collect demographic data about composers or performers as research subjects. Users should not treat synthesized timbres as biometric or identity-bearing human data. For concerns about a specific underlying score’s status, refer to PDMX’s documentation and applicable public-domain law; SPDMX inherits PDMX’s public-domain designations rather than conducting a new rights clearance for each item. Rendering uses redistributable soundfonts and open neural synthesizer weights (Apache 2.0; MIDI-DDSP / DDSP-Piano), not closed commercial VST sample libraries. Neural models were trained upstream on purpose-recorded research corpora (e.g. URMP, MAESTRO); SPDMX emits new synthesized audio conditioned on public-domain MIDI.
Intended Use
This dataset is intended for machine learning and MIR research and development, including source separation, multi-instrument transcription, generative music, stem-aware retrieval and data augmentation, and related audio/symbolic modeling. It is synthesized training/evaluation audio paired with MIDI, not a singer-identity corpus.
Provenance. SPDMX is a listening-guided rendering of PDMX, a large public-domain MusicXML/MIDI collection.
song_id) and PDMX’s public-domain-oriented metadata as given.See third-party / ethics notes in the code repository for synthesizer provenance.
Preprocessing and layout.
chunk_N/ directories for sharding. On Mozilla Data Collective the corpus is two .tar.gz archives (parts 1 and 2). Extract both into one directory. Even chunk ids are in part 1; odd chunk ids are in part 2. chunk_assignment.csv keeps each song in the same chunk across later releases.Scale. 253,982 songs; 490,159 mono stems; ~5,469 mixture hours @ 44.1 kHz. Demo and paper materials: pnlong.github.io/SPDMX
Disclaimer. Provided AS IS without warranty of title, non-infringement, or fitness for a particular purpose, to the extent permitted by law. CC BY 4.0 terms apply to the SPDMX release files themselves.