Demucs, online

Meta's four-part separation model, without the command line or a local GPU.

Drop your song here

or click to choose · mp3, wav, flac, m4a · up to 50 MB

What do you want?

VocalsTake the singing out, or keep only the singing.You get 2 files:VocalsSong without vocalsDrumsGet the drums alone, or play along without them.You get 2 files:DrumsSong without drumsBassGet the bassline alone, or play along without it.You get 2 files:BassSong without bassEverythingEvery part on its own, and the song minus each part.You get 8 files:VocalsDrumsBassOther (guitars, keys, synths)Song without vocalsSong without drumsSong without bassSong without the rest

Demucs is the open-source separation model developed at Meta, and for four-way splitting it is the reference implementation most other tools are measured against. Running it yourself means Python, pip, a model download and ideally a CUDA-capable GPU.

This runs it for you. Choosing Everything, Drums or Bass here executes htdemucs — Demucs v4 — on a hosted GPU and hands you the output.

What Demucs does that a two-stem model does not

A vocals-only separator answers one question: which parts of this mix are singing? Demucs answers four at once, dividing the recording into vocals, drums, bass and everything else. That extra structure is what makes drumless practice tracks, isolated basslines and full stem sets possible.

It costs more compute for the same song, which is why the vocals-only path here uses a lighter specialist model instead. If all you want is an instrumental, that path is both faster and marginally cleaner at the specific job.

The version running here is htdemucs, the hybrid transformer variant, which is the current default in the Demucs project and the strongest of the standard four-stem options.

Practical differences from running it yourself

No install: no Python environment, no torch version conflicts, no model checkpoint to fetch.
No GPU needed on your side. Demucs on a CPU is slow enough — several minutes for one song — that most people give up on it.
No command-line flags. Shifts, overlap and segment settings are fixed at values chosen for a sensible speed-to-quality balance.
No batch processing, and a five-a-day limit per device. Running it locally has no such cap, which is a real advantage if you have a lot of files.

Who ends up looking for this

People who tried the install and stopped. a dependency conflict or a missing CUDA toolkit is where most attempts to run it locally end, and the fix is rarely worth it for one song

Anyone on a locked-down machine. work and school computers where installing a Python environment is not an option, but a browser tab is

Users on Apple silicon or an old laptop. it runs on CPU, but slowly enough that a full song becomes a coffee break rather than a minute

Questions

›Which Demucs version is this?

htdemucs, the hybrid transformer model — Demucs v4, the project's current default four-stem model.

›Do I get the same output as running it locally?

Very close, though not necessarily bit-identical. Inference settings such as the number of random shifts affect the result slightly, and those are fixed here rather than exposed.

›Can I get the six-stem version with guitar and piano?

Not currently. A six-source Demucs variant exists that attempts guitar and piano separately, but it is meaningfully less accurate than the four-stem model, so it would need to be an additional option rather than the default.

›Is this affiliated with Meta or the Demucs authors?

No. Demucs is their open-source work, released under a permissive licence; this is an independent service that runs it. All credit for the model belongs to its authors.

Where it struggles

You get one model with fixed settings, which is the trade for not maintaining an install. Anyone who wants to experiment with shifts, overlap, segment size or alternative checkpoints will be better served running Demucs locally, where all of that is available and there is no daily cap.