StemDeck

If you like StemDeck and find it useful, consider buying me a coffee ☕.

Paste a YouTube URL, get the audio split into stems (vocals / drums / bass / guitar / piano / other) and play them back in a DAW-style multitrack mixer. Mute, solo, mix, zoom the waveform, loop a region, and download individual stems or a custom mix.

Local-only, single-user. One Python process serves both the REST/SSE API and the static frontend.

What is this? StemDeck is meant to be an open, free, simplified alternative to commercial stem-splitter products like Moises, LALAL.AI, and similar tools. It runs entirely on your own machine, with no account, no quota, no upload, and no subscription. Source-available so you can read, fork, or self-host it. It is not trying to match those products feature-for-feature; it covers the basics well and stops there.

Honest comparison

	StemDeck	Moises / LALAL.AI / similar
Price	Free	Freemium or pay-per-use
Hosting	Runs locally on your machine	Cloud, upload required
Account / login	None	Required
Stem model	Demucs htdemucs_6s (6 stems)	Proprietary, often higher quality
Polish	Functional, hobby-grade UI	Polished apps (web + mobile)
Extra features (pitch shift, chord detection, click track, lyrics, BPM tap, mobile apps, cloud library)	No	Yes
Privacy	Audio never leaves your machine	Audio uploaded to a third party
Source code	Available, modifiable	Closed

If you need the polish, the mobile app, or the extra musician-facing tooling, the commercial products are a better fit and worth the money. If you mainly want stems for personal study and prefer to keep things local and free, StemDeck should be enough.

Features

• 6-stem separation via Demucs htdemucs_6s. Auto-detects the best Torch device (CUDA, MPS, CPU); on Apple Silicon you get ~3-5× speedup over CPU for free.
• DAW-style waveform editor. Min/max sample rendering across all stems with shared global normalization, zoom in / out / Fit (+ / − / Cmd-wheel), loop drag on the ruler, gold playhead overlay, and stem-aligned lanes.
• Stem subset extraction. Click stem chips on the import page to pick which stems to keep. Filter-chip semantics: clicking from "all selected" snaps to "only this one"; subsequent clicks add or remove. Selection persists in localStorage.
• "Original" backing track. When you pick a subset, the studio includes a 7th lane with the complement (full song minus the selected stems). Playing it alongside the isolated stems reconstructs the full mix without doubling, which is perfect for A/B reference.
• Downloadable selected mix. A single mix.wav of just the selected stems, summed via ffmpeg amix. Surfaces as the Download Mix button in the footer.
• Per-stem mixer. Volume fader, mute, solo, and "monitor" (solo-only) per stem. State is synced between the preview mixer and the stems sidebar.
• Live VU per stem. Post-gain RMS via Web Audio analysers on each stem's gain node. Peak hold + slow falloff for the classic DAW meter feel.
• Song analysis. BPM (librosa beat tracker on percussive HPSS), key + scale + confidence (Albrecht-Shanahan profile with root-prominence weighting), integrated LUFS (BS.1770 via pyloudnorm), sample peak in dBFS. All surfaced in the now-playing card.
• Cancellable jobs. Click cancel mid-pipeline (download, Demucs, ffmpeg amix) and the runner terminates the active subprocess immediately, deletes the partial job dir, and returns the API to ready.

Requirements

• Python 3.10+
• ffmpeg on PATH (install instructions per-platform below)
• uv (recommended) or pip
• ~170 MB free disk for the Demucs htdemucs_6s model (downloaded automatically on first run)
• Reasonably modern CPU. An Apple-Silicon mps or NVIDIA cuda GPU dramatically speeds up separation.

Setup

macOS / Linux (one-shot)

git clone <repo-url> stemdeck && cd stemdeck
./run.sh setup     # detects OS, installs ffmpeg + uv, runs `uv sync`
./run.sh start

setup uses Homebrew on macOS and apt-get on Debian/Ubuntu. It skips anything that's already installed. For other Linux distros, install ffmpeg and uv manually, then run uv sync followed by ./run.sh start.

Open http://localhost:8000.

Running

run.sh (macOS / Linux)

The bundled control script manages the dev server as a background process with a PID file and a log:

./run.sh setup      # one-shot: install ffmpeg + uv (brew/apt), then `uv sync`
./run.sh start      # boots uvicorn, writes .run/uvicorn.pid + .run/uvicorn.log
./run.sh stop       # graceful shutdown, force-kills after 5 s, sweeps stray demucs children
./run.sh restart    # stop + start
./run.sh status     # is it running, on which URL

Environment variables it respects:

Variable	Default	Purpose
HOST	127.0.0.1	Bind address. Set to 0.0.0.0 to expose on the LAN.
PORT	8000	Listening port.
RELOAD	0	Set to 1 to pass --reload for hot reload during development.
FOREGROUND	0	Set to 1 to run uvicorn in the foreground (blocking, no PID file dance).

Examples:

RELOAD=1 ./run.sh start          # dev mode with auto-reload on file change
PORT=9000 ./run.sh start         # listen on a different port
HOST=0.0.0.0 ./run.sh start      # accessible from other machines on the LAN
FOREGROUND=1 ./run.sh start      # blocking; Ctrl-C to stop
tail -f .run/uvicorn.log         # follow logs while running in background

Manual uvicorn (any platform)

uv run uvicorn app.main:app --reload
# or:  source .venv/bin/activate && uvicorn app.main:app --reload

Docker

The Dockerfile and compose file live in build/:

docker compose -f build/docker-compose.yml up --build

Open http://localhost:8000. Stems land in ./jobs/ on the host. Demucs model weights are kept in a named volume (stemdeck-cache) so they don't re-download on rebuild.

Caveats:

• No GPU on macOS Docker. Docker Desktop on Mac doesn't expose MPS or CUDA, so Demucs runs CPU-only inside the container, which is significantly slower than the native install. Use the native install on Apple Silicon for speed.
• NVIDIA GPU passthrough (Linux) isn't enabled in build/docker-compose.yml by default. Add deploy.resources.reservations.devices with driver: nvidia if you want it; you'll also need nvidia-container-toolkit on the host.
• First build is slow. pip install . pulls torch (~2 GB) and friends. Subsequent rebuilds reuse the layer unless pyproject.toml changes.
• First job is slow. Demucs downloads htdemucs_6s weights (~170 MB) on the first run; cached after that.

How to use

1. (Optional) On the import page, click stem chips to pick which stems to extract. By default all 6 are selected. Clicking a chip while everything is highlighted snaps to "only this stem"; subsequent clicks add to or remove from the selection.
2. Paste a YouTube URL and click Process.
3. Wait through Downloading…, Analyzing…, Separating…, Mixing tracks…. First run also downloads the Demucs model (~170 MB).
4. When it's done, the studio dashboard appears. If you picked a subset, the first lane is Original (the song minus the selected stems); the rest are the isolated stems you chose. Otherwise all 6 stems show.
5. Mix:
   • ▶ ⏸ ⏹ master transport
   • M mute, S solo (additive: solo two stems and both stay audible)
   • ○ Monitor solo-only this stem (clears other solos)
   • vol fader: drag for 1:1 movement; double-click resets to 0 dB; Shift+wheel for coarse, plain wheel for fine adjustment
   • Reset / Mute / Solo toolbar buttons at the top reset/toggle all stems at once
6. Wave editor: drag on the ruler to define a loop region, click Loop to enable, + / − / Fit to zoom, Cmd/Ctrl-wheel to zoom centered on cursor.
7. Download Mix in the footer gives you a single WAV of just your selected stems summed together.

Keyboard shortcuts: Space play/pause · [ seek -5s · ] seek +5s · L toggle loop · I set loop start at playhead · O set loop end at playhead.

Layout on disk

jobs/<job_id>/
└── stems/
    ├── vocals.wav      # always: the 6 demucs stems
    ├── drums.wav
    ├── bass.wav
    ├── guitar.wav
    ├── piano.wav
    ├── other.wav
    ├── original.wav    # only when a strict subset was selected:
    │                   #   sum of the un-selected stems
    └── mix.wav         # only when ≥2 stems were selected:
                        #   ffmpeg amix of the selected stems

The source download is deleted after Demucs finishes (it's 100-300 MB and isn't used again).

Job state is in-memory only. Restart the server and the job list is wiped, but the files persist on disk. Send DELETE /api/jobs/<job_id> (or just delete jobs/<job_id>/) to reclaim space. Old job dirs are also swept automatically when a new job is submitted (TTL 24 h, configurable via STEMDECK_JOB_TTL_SECONDS).

Configuration

App environment variables, all optional:

Variable	Default	Purpose
STEMDECK_DEMUCS_DEVICE	auto	Force a Torch device: cuda, mps, or cpu.
STEMDECK_DEMUCS_MODEL	htdemucs_6s	Demucs model name.
STEMDECK_JOBS_DIR	./jobs	Where job dirs land.
STEMDECK_MAX_DURATION_SEC	1200	Reject videos longer than this.
STEMDECK_JOB_TTL_SECONDS	86400	How long to keep a job dir on disk.
STEMDECK_MAX_PENDING_JOBS	3	Admission control (currently advisory).

Troubleshooting

• uvicorn not found at .venv/bin/uvicorn when running ./run.sh start: you haven't installed deps yet. Run uv sync first.
• ffmpeg: command not found during a job: install ffmpeg per the platform-specific Setup section above and restart the server (./run.sh restart).
• WARNING: [youtube] No supported JavaScript runtime could be found: yt-dlp needs a JS runtime to reliably pick the best YouTube audio format. Install deno (brew install deno on macOS, curl -fsSL https://deno.land/install.sh | sh on Linux) and restart. Without it, downloads still work but may pick suboptimal formats.
• First separation is very slow: the Demucs model weights download on first run. Subsequent runs reuse the cached weights in ~/.cache/torch/hub/checkpoints.
• Demucs runs on CPU and takes minutes: Demucs picks the best device automatically (CUDA, MPS, CPU). On Apple Silicon you should see MPS acceleration; if not, your torch install may be CPU-only. Check the server log on startup for demucs config: model=htdemucs_6s device=mps.
• Browser memory grows on long videos: the multitrack player decodes each stem into a Web Audio buffer for the overview waveform. A 6-minute song uses a few hundred MB of decoded audio in memory; long lectures will be uncomfortable. Trim the input or close other tabs.
• Page reloaded mid-job: the job keeps running on the server, but the UI loses track of it. Wait for it to finish, then re-submit (the stems on disk get overwritten, or, if you want the previous output, find it under jobs/<job_id>/stems/).
• ./run.sh: Permission denied: the script lost its executable bit. Run chmod +x run.sh.

API (for tinkering)

Method	Path	Purpose
POST	/api/jobs	Body {url, stems?} → {job_id}. stems is an optional array of stem names; defaults to all 6.
GET	/api/jobs/{id}	Job state snapshot.
GET	/api/jobs/{id}/events	Server-Sent Events stream of job state.
POST	/api/jobs/{id}/cancel	Set the cancel flag and terminate the active subprocess.
GET	/api/jobs/{id}/stems/{name}.wav	Stream/download a single stem (range requests). name ∈ {6 demucs stems, original, mix}.
DELETE	/api/jobs/{id}	Remove the job dir from disk (must be done/error/cancelled).

Disclaimer

StemDeck is an educational project intended for study, research, and experimentation with audio source separation. It bundles two notable third-party tools:

• yt-dlp for downloading audio from URLs you provide.
• Demucs (htdemucs_6s model) for source separation.

You, the user running this software, are solely responsible for how you use it, including:

• Complying with the terms of service of any site you download from (YouTube's ToS, in particular, restricts automated downloads).
• Respecting the copyright of the audio you process. Separating stems from copyrighted material may be permissible for personal study or fair-use analysis in some jurisdictions, but redistribution of the resulting stems generally is not.
• Following the licenses of the underlying tools and models (yt-dlp, Demucs, FFmpeg, PyTorch, etc.).

The author(s) of StemDeck provide this code "as is", without warranty of any kind, and accept no responsibility or liability for how it is used. If you're unsure whether a particular use is allowed in your jurisdiction, consult a lawyer before proceeding.