Split stems on your own machine: a Demucs guide, and why unreleased tracks should stay off web uploaders

The quickest way to pull a vocal out of a mix is to drag the file onto a free website. For a song that's already out, that's mostly a rights question. For your own unreleased mix, or a client's session, it's also a custody question: once the file is on someone else's server, how long it stays there is up to them.

Demucs gives you another option. It's an open-source separation model, released under the MIT licence, that runs on your own computer. Here's how to set it up, what to expect from it, and when it's the right tool.

Is Demucs still maintained?

Only just, and you should know that before you rely on it. Meta's original repository was archived on 1 Jan 2025 and is now read-only. Its README says plainly that it "is not maintained anymore", because its author, Alexandre Défossez, no longer works at Meta.

He moved the project to his own fork, which he calls "the officially maintained Demucs". He adds that he isn't actively working on it, so expect slow replies and no new features. The PyPI package now points to that fork as its homepage. Version 4.1.0 was uploaded on 11 July 2026 and needs Python 3.10 or later.

In practice, follow the fork's instructions, not the archived ones. The old README still says Python 3.8 and a plain pip install. The fork has moved to the uv tool, which installs Python for you.

How do you run it?

You need a terminal, a few minutes, and a WAV, FLAC or MP3 file.

1. Install uv. On a Mac, the macOS guide suggests brew install uv. On Windows, the Windows guide gives winget install --id=astral-sh.uv -e, run from a command prompt. The Windows guide also says Demucs doesn't run on 32-bit systems.
2. Optionally install ffmpeg. The fork's README says you only need it for FLAC output and for less common input formats. WAV, MP3 and FLAC are read without it.
3. Run the command. Type uvx demucs, add a space, then drag your audio file into the terminal window to paste its path. The Windows guide suggests the drag trick, and says you need quotes around the path if it contains spaces.
4. Wait for the first run. Demucs needs its model weights before it can split anything. The model-loading code fetches them from the Hugging Face hub and falls back to Meta's download server. That is the network traffic. By our reading of the code, your audio isn't part of it: the split runs on your machine.
5. Find your stems. They land in a folder called separated/htdemucs/ followed by your track's name. Inside are four stereo WAV files at 44.1 kHz: drums.wav, bass.wav, other.wav and vocals.wav, per the README.

A few options worth knowing, all from the fork's README:

• --two-stems=vocals gives you a vocal and everything else (a karaoke split). It doesn't run any faster, because Demucs separates fully and then mixes the stems back together.
• --int24 or --float32 saves higher bit-depth files. The default is 16-bit, which is a poor choice if you're going to process the stems further.
• -n htdemucs_ft uses a fine-tuned model that takes four times as long and "might be a bit better".
• -n htdemucs_6s adds guitar and piano stems, but the README warns the piano stem "is not working great".
• -d cpu forces the processor if your graphics card runs out of memory. On Apple Silicon, the macOS guide says the GPU is used automatically. Intel Macs need uvx --python 3.12 demucs.

If the terminal isn't for you, the README links to third-party front ends, a desktop GUI and Ultimate Vocal Remover. We haven't audited either, so check their own documentation for what they send online.

What do you give up?

Fewer stems, slower runs and older benchmarks.

Demucs gives four stems by default. The best figure in the README, 9.0 dB signal-to-distortion ratio (SDR) for the fine-tuned model, comes from 2022 research on the MUSDB HQ academic test set, with training on that set plus 800 extra songs. That's a lab score, not a listening test, and it tells you little about a dense modern mix. We haven't run our own comparison.

Speed depends on your hardware. Without a usable GPU, the README says processing takes roughly 1.5 times the length of the track. A GPU needs at least 3 GB of memory, and about 7 GB with the default settings. One trap: when separation artefacts would make a stem clip, Demucs turns each stem down automatically. That can upset the balance between stems. --clip-mode clamp switches the auto-rescaling off.

Commercial tools have moved local too. We covered LALAL.AI's Lyra plugin, which splits locally at a quality the company itself puts below its cloud models, and Ableton Live's built-in stem separation, which also runs on your computer. Demucs is the free option that doesn't need a subscription or a Suite licence.

Why not just use a website?

Because a policy is a promise, and local processing doesn't need one. LALAL.AI is one of the better-documented services. Its privacy policy (updated 30 July 2026) lists a "history of processed files including physical files" among the data it collects. It says it keeps personal data "only for the period necessary", with no fixed number of days. It also says it doesn't use user files for AI training. That is a clearer answer than many tools give, but it is still a promise about a copy that sits on someone else's server.

Even Demucs has cloud routes. The archived README points to a Google Colab notebook and to free online services that run Demucs. Colab is, in Google's words, "a hosted Jupyter Notebook service". So your file goes up to a cloud virtual machine, not your own computer. An open model doesn't make someone else's server yours.

Before you upload anything unreleased, check:

• Is this my audio, or am I bound by a client agreement or an NDA?
• Does the service give a retention period in days, in writing?
• Does it say whether uploads are used to train models?
• Can I delete the file and the result myself?
• Would a local tool (Demucs, a DAW's built-in splitter) do the job well enough?

If any of the first four answers is no or unclear, split the file locally.

One more thing: separating a stem doesn't give you any rights to it. Pulling a vocal from someone else's record for a remix or a sample still needs clearance. This is general information, not legal advice.

What we think

For demos, rehearsal tracks, practice loops and remix sketches, local open-source separation is good enough, and it should be the default for anything unreleased. The privacy case is about how the tool is built, not about trusting a policy. If the audio never leaves your machine, there's no retention clause to read and no third-party copy to worry about.

The real risk with Demucs is that nobody is building it any more. In our reading, it rests on one maintainer who says he isn't actively working on it, on top of PyTorch, which moves on without it. The Intel Mac workaround already shows how a dependency can leave older hardware behind. Keep a working install, write down the version that works for you, and don't build a business workflow on it without a plan B.

The trend is good for artists. Splitting used to be a server-side service sold by the minute. It's becoming something that runs on your laptop or inside your DAW, and that leaves free web uploaders competing mainly on convenience. Local-first is worth rewarding wherever it shows up, and the same question applies to DAWs that now build stem separation in, Veena among them: ask where the processing runs.