Stems · measured 17 Oct 2026
Stems Terms signed
Every lane of a finished song, printed by the producer from the session it was made in. One 24-bit WAV per lane, MIDI where the session had it. Pop, house, hip hop and more.
No separation model touched these files. Each WAV is one lane of the producer's session, printed from the original project with its plugins loaded. Each lane carries one of 53 instrument labels in the manifest, taken from the producer's own lane name.
Every delivery includes the unmastered mix of each song, summed from its lanes, so a separation model learns from the source the parts came from, not from a guess. Train on these pairs and it pulls cleaner parts out of any finished mix. A mastered reference of the same song on request.
Inside the numbers
Sample rates, by pack
Genres
Lanes per song
Median 22 lanes. Eight songs in ten have between 12 and 41.
Instruments, by lane
Counted 16 October 2026 over 77,323 lane files in 2,770 stems packs on the store, printed songs and producer packs together: 69% labelled from the producer's own lane name, 27% by an audio classifier, 4% unlabelled. All 53 labels are listed. Lane counts are from the 2,315 songs printed from sessions.
Tempo
63 to 195 BPM, median 125.
Keys
1,247 major, 1,271 minor.
Formats
WAV 24-bit stereo, AIFF, MIDI
What a delivery contains
One folder per song, one WAV per lane with the producer's lane name, the unmastered mix of the song, a MIDI folder where the session had notes, and the manifest row with one instrument label per lane.
Manifest fields: song_id, creator_id, daw, bpm, key, sample_rate, stems, lane_label, duration, has_midi.
| song_id | creator_id | daw | bpm | key | sample_rate | stems | duration |
|---|---|---|---|---|---|---|---|
| s-004a1f | c-91b | Logic | 126 | F min | 48000 | 27 | 5:02 |
| s-004a20 | c-91b | Logic | 124 | A min | 44100 | 22 | 4:31 |
| s-004a2c | c-0e7 | Ableton | 128 | G min | 44100 | 19 | 3:58 |
What ships with every delivery
- One folder per song in the MoisesDB layout, readable by the moisesdb loader as is: one WAV per lane under its stem group, with the producer's lane name.
- mix.wav, the unity-gain sum of the lanes as 32-bit float, with the residual of a second sum recorded per song, and mix_flat_24.wav, the same sum at -0.3 dBFS peak in 24-bit with its gain recorded.
- data.json per song: DAW, BPM, key, sample rate, duration, and per lane the instrument label, peak and activity ratio, with has_bleed false by construction.
- A MIDI folder where the session had notes, and the session's own main render where it had one.
- checksums.sha256 in every song folder and SHA256SUMS for the whole delivery.
- manifest.csv and manifest.jsonl, one row per song, and index.csv listing every song in the catalogue.
- splits.json, a creator-level train, val and test split that holds across deliveries, and groups.json, the DAW versions of one song.
- DATASHEET.md: sources, dates, export settings, processing, de-duplication, exclusions, splits, known limits.
Rights for this dataset
Every creator in this set accepted terms that name AI training. Every stem is printed from the creator's own session, part by part, and the manifest says which session it came from. Read the rights page.
Often requested with
- DAW sessions The producer's project file with its audio and MIDI, as it was when the track was finished.
- MIDI files Note data from sessions and MIDI packs: chords, leads, basslines, drums.
- Vocal and instrumental pairs The vocal and the rest of the song as two aligned files from the same session.