Demucs online, in your browser
Extractune runs HTDemucs, the model Meta's Demucs v4 uses by default, inside your browser.
No file open
Stays on this device
Drop a file here
Audio or video, up to 500MB, 20 min
Vocal remover and stem splitter
No Python, no install, no GPU setup. You get the same four parts: vocals, drums, bass and other.
How it works
- 01
Drop the song here
The song itself (MP3, WAV, M4A, FLAC) or a video with it in, up to 500MB, on this page.
- 02
It splits on its own
The first split downloads the model to your browser once (about 90MB); after that it's kept. HTDemucs then runs on your graphics chip where the browser allows it, or on the processor.
- 03
Download the parts
The mixer saves each part as MP3, M4A, WAV, FLAC or WebM, all four as a zip, or a mix of your own; the Practice tab slows it down or moves it to another key.
Up to 6 minutes per split with a graphics chip, 2 without. Longer songs are split in 30-second batches that overlap, so there are no seams.
What people use it for
- Get Demucs stems without installing Python or a GPU driver.
- Split a song on a laptop or phone that can't run Demucs locally.
- Make an acapella or an instrumental for a remix.
- Practice with the drums or bass taken out.
Demucs in the browser vs on your computer
The model is the same; the setup isn't. Demucs itself is a Python program: per its README, you install it with pip and run it on a file from the command line, and it saves the parts as WAV files in a folder of their own. That gets you every model in the project, at the price of a Python setup and, for speed, a graphics card with a few gigabytes of memory.
This page runs htdemucs, Demucs's default model, converted for the browser. Our copy stores the weights in half the space, about 90MB to download once, and computes the same result as the original file. A long song goes through in 30-second batches, and inside them the model's 7.8-second windows overlap by 10% and are blended across each join. You can listen to the first batches while the rest are still splitting.
How long it takes depends on your device. Averaged over visitors' own splits, a 3-minute song takes about 1 minute on a computer with a graphics chip and about 4½ minutes on a phone with one. On the computer we test on, using WebGPU, the 3:18 test song took 27 s on 6 November 2026, 7.39 times faster than it plays.
Pick Python Demucs for the fine-tuned or six-part models, for batches of whole albums, or for files longer than a split here. Pick this page when you want the four parts now, on a laptop or a phone, with nothing to install.
| Detail | Demucs on your computer | On this page |
|---|---|---|
| Setup | Python, then python3 -m pip install -U demucs | None: a browser tab |
| Models | htdemucs (the default), htdemucs_ft, htdemucs_6s, hdemucs_mmi and the mdx family | htdemucs |
| Parts | Four; six with htdemucs_6s; two with --two-stems | Four: vocals, drums, bass and other |
| With a graphics chip | At least 3GB of GPU memory, about 7GB with the default settings | Any chip the browser can use through WebGPU |
| Without one | On the processor, in roughly 1.5 times the track's length | On the processor, up to 2 minutes a split |
| Length | Whole files; --segment trades quality for less memory | 6 minutes per split (5 with less memory) |
| Saves | 16-bit WAV by default, 24-bit or float WAV, or MP3 at 320kbps | MP3, M4A, WAV, FLAC or WebM, one part, all four as a zip, or your own mix |
Demucs alternatives compared: Hugging Face, Colab, a desktop app, Extractune
Search for Demucs online and you mostly find four kinds of thing: a demo someone hosts on Hugging Face, a Google Colab notebook, a web app that wants you to sign in, or the program on your own computer. They run the same family of models, but the song goes to different places.
The table describes each kind in general terms. Individual demos differ, so read the page of the one you pick before you upload anything.
| Way | What it is | Sign-in needed? | Where it runs | Your song |
|---|---|---|---|---|
| A Hugging Face Space | A hosted demo someone published | Varies by Space | On the Space's server | Uploaded to it |
| A Colab notebook | A notebook that installs and runs Demucs for you | A Google account to run it | On a Google machine | Uploaded to your Colab session |
| A web app with accounts | A site that wraps Demucs, such as demucs.danielfrg.com, which says to sign in with Google to upload files | Yes | On the site's server | Uploaded to it |
| Demucs on your computer | The Python program, from the command line or through a desktop app built on it | No | On your computer | Stays on it |
| This page | HTDemucs converted to run in the browser | No | In your browser tab | Stays on your device |
htdemucs, htdemucs_ft and v4: which one runs here
Demucs v4 is the version number of the project; htdemucs, the Hybrid Transformer Demucs, is the model v4 uses by default. Per the Demucs README, htdemucs_ft is a fine-tuned copy of it that takes about four times as long for slightly better results, and htdemucs_6s adds guitar and piano as parts.
This page runs htdemucs, so you get four parts: vocals, drums, bass and other. If you need the fine-tuned model or the six-part one, you need Demucs on your computer or a service that offers them.
Questions
Is this the real Demucs?
It's HTDemucs, the hybrid transformer model that Demucs v4 uses by default, converted to run in a browser. Same model, same four parts (vocals, drums, bass, other). It isn't the fine-tuned htdemucs_ft or the six-part version.
Do I need Python or a GPU?
No. It runs in Chrome, Edge or Safari. With a graphics chip the browser can use (WebGPU), it's much faster; without one it runs on the processor, for shorter clips.
Is my song uploaded?
No. The model is downloaded to your browser and the split runs on your device. Your file never leaves it.
Why is the first split slow to start?
The model (about 90MB) is downloaded once, the first time. Your browser keeps it, so later splits start straight away. Wi-Fi is recommended for that first time.
Is it free?
Yes. No sign-up and no watermark. The split runs on your device, so there's no monthly count.
More practice tracks
Updated