Vocal remover & stem splitter
Remove or isolate vocals from a song with AI — split it into vocals, instrumental, drums, bass and other. Private — runs on your device
Powered by Meta's open-source Demucs model running in your browser. checking…
Click to choose or drag & drop a song here
MP3, WAV, M4A/AAC, OGG, FLAC · one track at a time
Notes
How it works: the first time you separate a track, an AI model (~172 MB) downloads to your browser and is cached — later runs skip the download. Separation then happens entirely on your device: on your GPU via WebGPU where available, otherwise on the CPU (slower). A typical song takes from about a minute to several minutes depending on your hardware.
Stems you get: Vocals and Instrumental (the whole backing track), plus Drums, Bass and Other on their own — all as lossless WAV. Nothing is uploaded; your audio never leaves your device.
Heads-up: long tracks use a lot of memory — for very long files a desktop browser works best. Best results come from clear stereo mixes.
How to use the vocal remover
- Drop in a song — MP3, WAV, M4A/AAC, OGG or FLAC, one track at a time. The badge at the top shows whether it'll run on your GPU (WebGPU) or CPU before you start.
- Click Separate stems. The very first run downloads the ~172 MB Demucs AI model once; your browser caches it, so later runs start immediately.
- Watch it work: the track is processed in segments, with elapsed time, segment count, processing speed and an ETA shown live.
- Preview each stem in the built-in players, then download what you need: Vocals, Instrumental, Drums, Bass and Other — each a lossless WAV named after your file.
Common uses
- Make a karaoke or backing track: download Instrumental and the lead vocal is gone while drums, bass and everything else stay intact.
- Pull an a cappella (Vocals only) from a finished song for a remix, mashup or vocal study.
- Isolate Drums or Bass to learn a part by ear, build practice loops, or sample a groove.
- Solo each part of a mix you admire to hear how it was arranged and balanced.
Tips & limitations
- Expect the best results from clear studio stereo mixes. Live recordings, heavy reverb or dense walls of sound can leave audible bleed — faint traces of one instrument in another's stem.
- Stems come out as 16-bit stereo WAV at 44.1 kHz — about 10 MB per minute each. Run them through the audio converter if you'd rather have small MP3s.
- Long tracks are memory-hungry: the song plus five stems all sit uncompressed in RAM, so very long files can fail — especially on phones. A desktop browser is the safe choice.
- Speed depends on hardware: with WebGPU (a recent Chrome or Edge and a decent GPU) a typical song takes a minute or two; the multi-threaded CPU fallback works everywhere but can take several minutes.
How it's built & why it's safe
Uses Meta's open-source Demucs (Hybrid Transformer) source-separation model, run entirely in your browser with ONNX Runtime Web — on your GPU via WebGPU where available, otherwise multi-threaded WebAssembly across your CPU cores. Your audio is never uploaded; the only download is the model itself (~172 MB, fetched once from the Hugging Face model hub and cached). The AI predicts four sources — vocals, drums, bass and other — and the Instrumental is built by combining the last three. Because this on-device AI needs cross-origin isolation, which ad networks don't support, this page runs ad-free.
Related tools: Audio Converter · Video Converter · Image Converter
Further reading: How AI vocal removers actually work · Why Toolkit runs entirely in your browser (and why that matters)
Frequently asked questions
Is my music uploaded?
No. The AI runs entirely on your device, so your audio never leaves your browser. Only the model file is downloaded — once, from Hugging Face — and then cached for future visits.
How long does separation take?
It depends on your hardware and the track length. With a recent GPU (WebGPU) a typical song takes roughly a minute or two; on the multi-threaded CPU fallback it can take several minutes. Live stats show speed and an ETA while it runs.
Why is there a one-time download?
The Demucs AI model is about 172 MB — that's the neural network itself. It's fetched the first time you use the tool and cached by your browser, so later separations start with no re-download.
Which stems do I get?
Five: Vocals and Instrumental (drums, bass and other combined) as the two main outputs, plus Drums, Bass and Other individually — all as lossless 16-bit WAV files at 44.1 kHz.
Can it isolate guitar or piano separately?
Not on their own. The model separates a song into four sources — vocals, drums, bass and everything else — so guitars, pianos, synths and strings all land in the Other stem.
Is it really free?
Yes. It's an open-source AI model running on your own hardware, so there are no accounts, credits, per-song fees or upload limits — your computer does the work.
Why doesn't this page show ads?
On-device AI needs cross-origin isolation, which isn't compatible with our ad provider — so this one page is deliberately ad-free. The rest of the site is unaffected.