Hugging Face's logo

Datasets:
Maqsudjonpolatov
/
nullsample-multitrack

Tasks:
Audio Classification
Audio-to-Audio
Modalities:
Audio
Text
Formats:
soundfolder
Size:
< 1K
Tags:
music
Synthetic
procedural-audio
multitrack
stems
source-separation
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1
nullsample-multitrack / README.md
Maqsudjonpolatov's picture
Maqsudjonpolatov
Dataset card: 20 synthetic tracks with stems, MIDI and exact annotations
f1d5966 verified
preview code
|
Raw History Blame Contribute Delete
8.15 kB
metadata 
license: cc-by-4.0
pretty_name: Nullsample Multitrack
task_categories:
  - audio-classification
  - audio-to-audio
tags:
  - music
  - synthetic
  - procedural-audio
  - multitrack
  - stems
  - source-separation
  - music-transcription
  - drum-transcription
  - beat-tracking
  - chord-recognition
  - key-detection
  - midi
  - mir
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: mix/*

Nullsample Multitrack

20 complete tracks generated from code, with zero samples. Each comes with its mix, six stems, a MIDI file and exact annotations: tempo, downbeats, key, the chord in every bar, the sections, every drum hit by instrument and every note of every part. That is 40,060 drum hits and 32,919 notes over 37 minutes of music.

The labels are not estimated and not hand-annotated. They are the score the audio was synthesised from, so they are exact by construction. No recordings, samples or trained models were used. Every sound is computed from oscillators, noise, envelopes and filters by Nullsample (source, AGPL-3.0). Nothing in the audio came from anyone else's recording.

Every track can be reproduced bit for bit from its seed, on any machine. The engine avoids the platform-dependent Math.sin/exp/pow and is tested byte-identical across Node and browsers. You can listen to any seed at https://nullsample.maqsudjon.com/generate/#s=<seed>, or render the lossless WAV yourself:

git clone https://github.com/maqsudjon-cell/nullsample && cd nullsample && git checkout b6adf10
npm install && npm run render -- --seed NULL-0001 --out NULL-0001.wav
npm run dataset -- --out ./dataset        # this whole dataset

The audio came from engine commit c912de6 (engine_commit in every annotation file). The checkout above has the same engine plus the script that wrote this dataset.

What it is good for

  • Source separation: six stems per track (drums, 808 bass, lead, arp, pads, fx).
  • Drum and note transcription: onsets and notes that are exact to the sample.
  • Beat, downbeat and tempo tracking.
  • Key and chord recognition, and structure segmentation.
  • Testing an MIR tool against ground truth before trusting it on real recordings.

Files

Path What
mix/<seed>.mp3 the track, MP3 192 kbps, 44.1 kHz stereo
mix/metadata.csv one row per track: tempo, key, scale, bars, length, form, sections, chords. The viewer reads it.
stems/{drums,bass808,lead,arp,pads,fx}/<seed>.mp3 each part rendered on its own, MP3 128 kbps
annotations/<seed>.json the labels, times in seconds
midi/<seed>.mid the same notes as a Standard MIDI File

annotations/<seed>.json 

{
 "seed": "CLIP-N0ISE", "tempo_bpm": 152.5, "time_signature": "4/4", "bars": 68, "duration": 109.516,
 "downbeats": [0, 1.5738, 3.1475, ...],
 "key": { "tonic": "G", "scale": "phrygian", "scale_intervals": [0,1,3,5,7,8,10] },
 "form": "double-drop",
 "sections": [{ "name": "intro", "start": 0, "end": 6.2951, "intensity": 0.12 }, ...],
 "chords": [{ "start": 0, "end": 1.5738, "chord": "Gsus2", "root": "G", "voicing": "sus2", "notes": [55,57,62] }, ...],
 "drums": { "kick": [{ "time": 6.2951, "velocity": 1 }, ...],
            "hatClosed": [{ "time": 6.2951, "velocity": 0.825 }, ...],
            "hatOpen": [{ "time": 6.4918, "velocity": 0.629 }, ...],
            "rim": [{ "time": 6.4918, "velocity": 0.317 }, ...],
            "snare": [{ "time": 6.6885, "velocity": 0.995 }, ...],
            "clap": [{ "time": 7.4754, "velocity": 1 }, ...] },
 "notes": { "bass808": [{ "start": 5.9016, "end": 6, "midi": 24, "velocity": 0.686 }, ...],
            "lead": [{ "start": 6.2951, "end": 6.3934, "midi": 94, "velocity": 1 }, ...],
            "arp": [{ "start": 37.7703, "end": 37.8146, "midi": 110, "velocity": 0.95 }, ...],
            "pads": [{ "start": 0, "end": 1.3849, "midi": 67, "velocity": 0.8 }, ...] },
 "fx": [{ "kind": "riser", "start": 3.1475, "end": 6.2951 }, ...],
 "regenerate": "https://nullsample.maqsudjon.com/generate/#s=CLIP-N0ISE",
 "engine_commit": "c912de6da8dc5c66871a8acb8c0441a2d4666193"
}

The MIDI files have 480 ticks per beat and the track's tempo. They have a tempo track with the sections as markers, the drums on channel 10 (General MIDI: kick 36, rim 37, snare 38, clap 39, closed hat 42, open hat 46), and one track each for bass808, lead, arp and pads. Velocities in the JSON are 0–1. The pads are chords with no velocity of their own, written at 0.8. Chord names: 5 is a power chord, m minor, m7, sus2, sus4, add9, m(add9).

The tracks

Tempo 151–159.5 BPM. Scales: phrygian, minorPentatonic, naturalMinor, harmonicMinor, phrygianDominant.

Seed BPM Key Form Length Sections
CLIP-N0ISE 152.5 G phrygian double-drop 1:50 intro drop break drop2 break2 drop3 outro
DUCK-UNDER 155 F minorPentatonic double-drop 1:48 intro drop break drop2 break2 drop3 outro
FOLD-BACK3 159 C# minorPentatonic double-drop 1:45 intro drop break drop2 break2 drop3 outro
GLASS-EDGE 156.5 G naturalMinor cold-open 1:47 drop verse drop2 verse2 drop3 outro
GRID-LOCK4 159.5 A# harmonicMinor long-build 1:57 intro build drop break drop2 outro
HARD-CODE1 158 A naturalMinor double-drop 1:46 intro drop break drop2 break2 drop3 outro
K7M2-QX4B 154.5 D harmonicMinor long-break 1:42 intro build drop break drop2 outro
LAST-CALL9 154 B naturalMinor cold-open 1:48 drop verse drop2 verse2 drop3 outro
NULL-0001 155 E minorPentatonic sustained 1:54 intro hold lift hold2 lift2 outro
OSC-ILLATE 156.5 A harmonicMinor cold-open 1:47 drop verse drop2 verse2 drop3 outro
PHRY-GIAN9 156 E phrygian straight-in 1:53 drop break drop2 bridge drop3 outro
PRIME-DLAY 155.5 D harmonicMinor arc 1:54 intro build drop break drop2 bridge outro
RAGE-88KK 156.5 C# harmonicMinor sustained 1:53 intro hold lift hold2 lift2 outro
SAW-STACK5 155 E naturalMinor cold-open 1:48 drop verse drop2 verse2 drop3 outro
SEED-4242X 154 D# phrygian straight-in 1:55 drop break drop2 bridge drop3 outro
SINE-WAVE7 159.5 F# naturalMinor sustained 1:51 intro hold lift hold2 lift2 outro
SUB-DROP01 158 A harmonicMinor long-build 1:58 intro build drop break drop2 outro
TAPE-NULL2 151 F phrygianDominant long-break 1:44 intro build drop break drop2 outro
VOID-7X2A 151.5 C naturalMinor long-break 1:44 intro build drop break drop2 outro
ZERO-SAMPL 159 A phrygian straight-in 1:51 drop break drop2 bridge drop3 outro

Limits

  • One style. All tracks come from one preset (energetic electronic, hyperpop-like), so this is not a sample of music in general. It is small: use it to test and probe models, or as a clean supplement, not as a training corpus on its own.
  • Stems do not sum exactly to the mix. Each stem is its bus rendered alone through the master chain at a fixed drive. The mix also has mix-wide effects (tape stops, stutters) and a limiter that reacts to everything at once.
  • MP3, to keep it small. Annotation times come from the engine's sample positions. Decoders that honour the LAME gapless header, such as ffmpeg and most libraries, line up within a few milliseconds. For lossless audio, render the seed.
  • The synthesis parameters behind each sound are not included.

Licence

The audio, MIDI and annotations: CC BY 4.0. Please credit "Nullsample (nullsample.maqsudjon.com)". The engine that made them: AGPL-3.0. Audio you render yourself with it is yours.

@misc{nullsample_multitrack,
  title  = {Nullsample Multitrack: synthetic music with exact annotations},
  author = {Polatov, Maqsudjon},
  year   = {2026},
  url    = {https://nullsample.maqsudjon.com}
}