Drop in a metal, punk, or rock song → get the isolated drums and bass, a mixer to practice with, and auto-generated drum sheet music.
Everything runs locally — your audio never leaves your machine.
- Splits any song into stems (drums / bass / vocals / guitars) with Demucs v4, the open-source state of the art — on dense, distorted mixes included.
- Mixer-board web UI: a fader per stem with mute/solo, so you can drum along with the band minus the drummer, or isolate the kit to learn it.
- Practice tools: slow playback to 50% with pitch preserved, an A–B loop for drilling one riff or fill, and a hit-marker strip showing every detected drum hit on the timeline (click to jump).
- Drum sheet music: transcribes the drum stem to MIDI and engraves it — shown inline on the mixer page with the current measure highlighted as the song plays, downloadable as PDF/MusicXML/MIDI. Dynamics survive: ghost notes engrave parenthesized, accents get accent marks.
- Bass tabs: note transcription plus a playable ASCII tab (chord-aware — double-stops land on separate strings).
- Guitar tabs (optional): tick "also transcribe guitar" and the 6-stem model adds Guitar & Keys channels plus a separate chord-aware guitar tab. Honest caveat: clean/lead guitar transcribes usefully; high-gain rhythm guitar is the hardest signal there is — treat those tabs as a first sketch.
- Guitar Pro export: bass and guitar tabs also save as
.gp5(opens in Guitar Pro, TuxGuitar, MuseScore — with playback). - A real library: processed songs survive server restarts and are listed on the drop page; separation shows live percent progress while you wait.
- Download your mix: the stem levels you set render straight to a WAV in the browser.
| Drop a song in… | …then practice with it |
|---|---|
 |
 |
Everything above is real output. The demo song is a synthesized rendition
of the classic Smells Like Teen Spirit four-chord riff (F5–B♭5–A♭5–D♭5 at
117 BPM), generated by examples/demo_riff.py — run it yourself for an
instant test song plus every artifact the pipeline produces, no copyrighted
audio involved. The demo drums follow the transcribed groove: the kick pair
on 1 and the "&" of 1 into the backbeat, syncopated kicks on the "a" of 2 and
the "&" of 3, straight-8th hats (crash wash in the back half), and the
flammed snare/kick/hat entrance fill leading into the crashes.
This is the bass tab the pipeline wrote for that riff. Nobody told it the fingering — the chord-aware assigner lands on the classic shape by itself (F and B♭ at the 1st fret, A♭ and D♭ at the 4th):
== Smells Like Teen Spirit (demo riff) — bass ==
tuning: E1 A1 D2 G2 (low to high) tempo: ~117 BPM
G|-------------------------------------------------------------------------
D|-------------------------------------------------------------------------
A|-----1--1--1-----4--4--4--------1--1--1-----4--4--4-----1--1--1-----4--4-
E|--1-----------4-----------1--1-----------4-----------1-----------4-------
The same notes also land in bass.gp5 (opens in Guitar Pro / TuxGuitar,
with playback) and bass.mid. And this is the demo groove engraved by the
notation stage (from the demo's known hits — reference.musicxml; on a real
song the same engraver renders the transcriber's draft instead, which is
what shows inline on the mixer, synced to playback):
One line (macOS / Linux) — installs to ~/drum-extractor, pre-downloads
the model, launches the app, and opens your browser:
curl -fsSL https://raw.githubusercontent.com/Ben-K-Jordan/Drum-Extractor/main/bootstrap.sh | bashRelaunch any time with ~/drum-extractor/run.sh (macOS users also get a
double-clickable Drum Extractor.command).
Docker — no clone, no Python:
docker run -p 8237:8237 -v drumx-output:/app/output ghcr.io/ben-k-jordan/drum-extractorWindows — clone the repo, double-click install.bat, then run.bat.
Then drop a song in at http://127.0.0.1:8237 (the browser opens by itself outside Docker). Separation takes roughly the length of the track on CPU, much less with a GPU.
Manual install / picking your own pieces
git clone https://github.com/Ben-K-Jordan/Drum-Extractor.git && cd Drum-Extractor
./install.sh # scripted: venv + extras + launchers + checkup
# or fully by hand:
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[all]" # separation + transcription + notation + web UI
drum-extractor doctor # shows what's ready and how to fix any gapsPick-and-choose extras if you don't want everything: [separation] [drums]
[bass] [notation] [web] — each stage degrades gracefully when its
dependencies are absent, and error messages always name the exact install
command.
Optional upgrades
| What | Why | How |
|---|---|---|
| ADTOF | The best open drum transcriber (trained on rock/metal) — without it a rough built-in fallback is used | pip install -e ".[adtof]" |
| MuseScore 4 | PDF sheet export (MusicXML works without it) | musescore.org, CLI on PATH |
| madmom | Bar-accurate downbeat tracking (librosa fallback otherwise) | pip install -e ".[quantize]" — its old build chain can fail on modern Python, which is why it's not in [all] |
| audio-separator | Blend a second drum model with Demucs (--ensemble-model) |
pip install -e ".[ensemble]" — its strict dependency pins can fight the rest of the stack, which is why it's not in [all] |
Docker, built locally (instead of the prebuilt ghcr.io image):
docker build -t drum-extractor .
docker run -p 8237:8237 -v drumx-output:/app/output drum-extractorNot sure what's installed? drum-extractor doctor prints a checkup of
every feature with the exact fix for anything missing.
The web UI is the easy path; everything is also scriptable:
drum-extractor run song.mp3 # full pipeline -> output/song/
drum-extractor separate song.mp3 # stems only
drum-extractor transcribe-drums stems/drums.wav
drum-extractor notate output/song/transcription.json
drum-extractor run song.mp3 --grid 32 --boost-double-kick # fast-metal setup
drum-extractor run song.mp3 --guitar # + guitar tabs (6-stem model)
drum-extractor run song.mp3 --fixed-tempo 184 --grid-mode constant # you know the BPM
drum-extractor run song.mp3 --reuse-stems --drum-backend adtof # re-transcribe, skip separationoutput/<song>/ contains the stems, drums.mid, bass.mid, bass.tab.txt,
drums.musicxml (+ drums.pdf with MuseScore), a drums_sonified.wav you can
A/B against the stem by ear, and transcription.json (re-notate any time
without re-processing).
| Stage | Tool | License |
|---|---|---|
| Source separation | Demucs v4 htdemucs_ft |
MIT |
| Drum transcription | ADTOF (via ADTOF-pytorch), with a built-in onset fallback | non-commercial* |
| Bass transcription | basic-pitch + torchcrepe | Apache-2.0 / MIT |
| Beat tracking | madmom or librosa | BSD / ISC |
| Engraving | music21 → MusicXML → verovio SVG / MuseScore PDF | BSD / LGPL / GPL |
* ADTOF's weights are CC BY-NC — fine for personal practice; swap in OaF-Drums (Apache-2.0) if you need a commercial-safe chain. This project's own code is MIT.
Extras for fast metal: a double-kick booster (--boost-double-kick)
recovers blast-tempo kicks from the low band that full-kit models merge
(measured: kick recall 0.67 → 0.98 at 200 BPM), and an optional two-model
drum ensemble with UVR-style spectral blending cuts distorted-guitar bleed.
- Stem separation is the solved half. Drums come out clean even from dense metal mixes; bass is clearly usable, though down-tuned distorted bass loses some character.
- Sheet music is a draft, not gospel. A mid-tempo groove transcribes into
a genuinely useful chart. Fast metal — blast beats, 200+ BPM double-kick,
dense cymbal work — comes out as a scaffold you correct by ear; that's the
current limit of every transcription tool, commercial ones included. The
drums_sonified.wavrender exists exactly for that by-ear checking. - Cymbal/tom identity is the weakest area of automatic transcription; kick/snare/hi-hat are the reliable core.
The repo ships a ground-truth groove bank: programmatically generated grooves (d-beat, blasts, gallops, fills…) rendered by a built-in synthesizer, so the true hits are known and any transcriber can be scored with real F-measures:
drum-extractor bank-build -o bank
drum-extractor bank-eval bank --backend onset # or: --backend adtofCI runs the full test suite and this accuracy gate on every push — a change that degrades transcription fails the build. The tuning story (including an overfitting mistake the process itself caught) is in docs/development.md, along with the full QA history: every module was adversarially audited and diffed against its upstream project's real source.
./install.sh --light # no torch/TF stacks: drums, notation, web, gp5
.venv/bin/pytest -q # 150 tests, no GPU or model downloads neededIssues and PRs welcome — especially real-world reports of where transcription falls down on your songs, and hand-corrected charts (each becomes an (audio, ground-truth) pair the bank can score against).
Standing on the shoulders of Demucs, ADTOF, basic-pitch, torchcrepe, librosa, madmom, music21, verovio and MuseScore.