AI Vocal Remover — Remove Vocals in Your Browser, Free
Split any song into vocals and instrumental. Everything runs on your device.
63.7 MB model downloads once, then stays cached.
Drop an MP3 or WAV here, or click to browse
MP3 or WAV · unlimited runs · no length or size limit
Desktop only · WebGPU required · Free and unlimited · Audio never leaves your device
Three steps, no upload
Drop your track
MP3 or WAV, up to 5 minutes. Nothing is uploaded.
Separate on your device
A 67 MB model runs on your GPU. Later runs skip the download.
Download two MP3s
Instrumental and vocals, 192 kbps, ready for karaoke or remix.
Free Browser-Based Vocal Removal (No Upload)
The free tool loads an ONNX model into your browser and runs it on your GPU. There is no server round trip, no file upload, and no account.
You get two stems — instrumental and vocals — as 192 kbps MP3. There is no track length, file size, or run limit — the free tool is free every day, and the only requirement is a browser with WebGPU.
Works on Desktop with WebGPU
Chrome, Edge, and Safari 17+ run the model on your graphics card. No GPU? Cloud separation is one click away.
Your Audio Never Leaves Your Device
Local separation is pure browser code. There is no upload endpoint for it, so there is nothing to leak.
One 67 MB Model, Cached After First Run
UVR-MDX-NET-Voc_FT downloads once into your browser cache, then starts in seconds.
Model: UVR-MDX-NET-Voc_FT · MDX-Net architecture · runs via ONNX Runtime Web
Cloud Vocal Separation for Longer Tracks
Same model family, running on our GPUs. Upload once, get both stems back in about a minute — and handle tracks up to 15 minutes.
| Local (free) | Cloud | |
|---|---|---|
| Where it runs | Your browser (WebGPU) | Our servers |
| Track length | No limit (your machine decides) | 10 min (Starter) · 15 min (Creator) |
| Speed | Roughly 1× track length | Usually under a minute |
| Cost | Free, unlimited runs | From $2.99/month |
How AI Vocal Separation Works
What the Model Actually Does
MDX-Net estimates the vocal component of a spectrogram, then subtracts it from the original mix. The vocal stem is the estimate; the instrumental is the residual. That is why backing vocals and reverb tails can still be audible in the instrumental.
MP3 vs. WAV: What You Get
Both stems are encoded to 192 kbps MP3 at 44.1 kHz, which is plenty for karaoke, practice, and most remix work. Lossless WAV and 4-stem output are not part of this release.
Frequently Asked Questions
Is my audio uploaded anywhere?
With local separation, no: the model runs in your browser and the audio never leaves your device. Cloud separation does upload your file, and the output is deleted after 7 days while the input is removed as soon as the job finishes.
Why does local processing need WebGPU?
The separation model needs a GPU to run in reasonable time. If your browser can't provide WebGPU, we point you to cloud separation instead of running a slow CPU fallback.
Why is the first run slow?
The model file is about 67 MB and is downloaded once, then cached by your browser. Later runs start almost immediately.
What's the difference between vocals and instrumental?
They are complementary stems: the instrumental is the original mix minus the estimated vocal track. Backing vocals and reverb tails can bleed into the instrumental — that is a property of the model, not a bug.
Do cloud plans include commercial use?
Starter and Creator include commercial use of the stems you generate. The free browser tool is for personal use only. You are still responsible for having the rights to the original recording.
How are minutes counted?
We round the track length up to the next minute (3:20 counts as 4 minutes) and only charge successful jobs. Slow-queue jobs are never charged.