Stem Splitter
Split any track into vocals, drums, bass and more — free, on your own machine.
No upload. No account. No queue. No server ever hears your audio. The kind of privacy the cloud splitters structurally can't offer — because there is no cloud.
Download
SHA-256 — published with each release so you can verify your download:
0eb6455687ab8af2e0a625a5cd3919c1a4fd029160452100edb88c1743144d39
Windows will warn you the first time.
We're an indie build with no code-signing certificate yet, so SmartScreen says
"Windows protected your PC". Click More info → Run anyway.
That's the whole ritual — verify the SHA-256 above if you want certainty.
Quickstart
- Unzip anywhere you like (Desktop is fine — it's fully portable).
- Run
DMM Stem Splitter.exe. A server window opens, then your browser. - Drop in a track (or a whole EP), pick a mode, hit split. Stems land in your browser — audition them with the built-in mixer, download individually or as one zip.
First run: each AI model downloads its weights the first time you use it (~80–300 MB per model, once). After that, splitting works fully offline — airplane-mode-verifiable. Close the server window when you're done.
What it does
- 4-stem & 6-stem separation — vocals, drums, bass, guitar, piano, other
- Metal Hybrid mode — runs both models and assembles the best of each: the cleanest 6-stem split this engine can produce, built for dense metal
- Karaoke & backing tracks — isolate vocals, drums, bass or guitar; get Instrumental, Drumless, Bassless, Guitarless
- Quality presets — Standard / High / Ultra shift-averaging for dense mixes
- Real formats — WAV 16/24-bit, FLAC, MP3 320/256/192
- In-browser stem mixer — solo/mute/volume + waveforms before you download a thing
- Batch queue — drop an album before bed, wake up to stems
- Live progress + ETA — stage-aware, cancellable, survives a refresh
- Field Manual included — a Guide built into the app, plus a PDF manual in the zip
- Powered by Demucs — Meta AI research's separation model, running locally
FAQ
Is it really private?
Yes — structurally. The app is a local server on 127.0.0.1; your browser talks to your own machine. Your audio never leaves your PC — there is no upload anywhere, ever. The app touches the network for exactly one thing: the one-time weight download per model (from Meta's public checkpoint host). After that, the app makes zero network requests — even the fonts are baked in. Don't take our word for it: flip on airplane mode and it runs, looks, and splits exactly the same.
Why is the download bigger than a normal app?
There's a complete neural-network runtime (PyTorch) inside — the same machinery the cloud services run on their servers, except this copy is yours. ~200 MB zipped, ~600 MB unpacked, plus model weights on first use.
How long does a split take?
CPU-only by design (works on any PC): roughly 5–15 minutes for a 4-minute track on the default model, longer on the higher-quality presets. Queue a batch and let it run — that's what it's built for.
Where do my files go?
Your tracks and stems live in a data folder next to the exe (or %LOCALAPPDATA%\DMM Stem Splitter if the app folder isn't writable — e.g. under Program Files). Downloads go wherever your browser puts them. Model weights cache in %USERPROFILE%\.cache\torch so they survive app updates. No installer, no registry — delete the app folder plus that cache folder and every trace is gone.
Why does Windows warn me about the exe?
The build is unsigned (code-signing certificates cost real money; we'd rather keep this free). SmartScreen flags unsigned downloads on principle. More info → Run anyway, and verify the SHA-256 above if you want to be sure the file is untouched.
Does it run on anything besides 64-bit Windows?
64-bit Windows 10/11 covers effectively every Windows PC of the last decade. A 32-bit build isn't just unmade — it's impossible: the AI runtime (PyTorch) dropped 32-bit support years ago. Windows 11 on ARM (Surface and friends) runs it through the built-in x64 emulation — slower, but it works. macOS / Linux: not yet. A Mac build needs Apple hardware, signing and notarization, so it's gated on demand — if you want it, say so in the Discord.
What about GPU acceleration?
This build is CPU-only to stay small and run everywhere. A CUDA build roughly doubles the download — if enough people ask, we'll ship one.
The fine print
Split only audio you own or have the rights to work with.
Separation engine: Demucs (Meta AI research, MIT license).
Audio decoding: FFmpeg, bundled as separate programs — license texts ship inside the zip.
Built by Dark Matter Metal for the metal underground. 🤘