AI vocal remover. A real one, in your browser.
Separates the vocal from a song using a neural network that runs entirely on your machine — nothing is uploaded. You get a clean instrumental and an acapella. Unlike centre cancellation it works on mono files and on vocals that are not panned dead centre, because it recognises a voice rather than a position in the stereo field.
drop a song
MP3, WAV, M4A, FLAC. One track at a time, up to about ten minutes.
At a glance
| Model | MDX-Net (UVR-MDX-NET-Voc_FT) via ONNX Runtime Web |
|---|---|
| Outputs | Instrumental and acapella |
| Model download | 64 MB, once, then cached in your browser |
| Acceleration | WebGPU where available, CPU fallback otherwise |
| Speed | About the length of the track on a GPU; roughly ten times that on a CPU |
| Processing | Overlapping 5.9-second parts, cross-faded |
| Length limit | About ten minutes — both stems are held in memory |
| Uploads your file | No — only the model is fetched, and it holds none of your audio |
How this differs from the simple vocal remover
The vocal remover on this site cancels whatever is panned dead centre. That is arithmetic: instant, works anywhere, and effective on conventional stereo mixes where the lead sits in the middle. It also fails completely on a mono file, leaves a wash behind when the vocal has stereo reverb, and thins any instrument sharing the centre.
This page runs a model trained on thousands of songs to recognise what a voice looks like in a spectrogram. It does not care where anything is panned — feed it a mono recording and it still works, because it separates by timbre rather than by position.
The trade is time. Centre cancellation finishes before you look up. This downloads a neural network and then does real arithmetic on every frame of your song. Use the simple one when it works for your track; use this when it does not.
Why vocals and instrumental rather than four stems
Four-stem models — the ones that give you drums, bass, other and vocals separately — are a different weight class. The ONNX export of Demucs, the best known of them, is 158 MB, and ONNX Runtime Web cannot load it: session creation runs for two minutes inside the WebAssembly heap and then aborts. That is not a tuning problem, it is a memory ceiling.
MDX-Net is 64 MB, loads in about two seconds, and separates the one boundary most people actually want. It is the same family of model that Ultimate Vocal Remover uses on the desktop. Choosing the model that runs beat choosing the model with the better name.
What it gets wrong
Separation is very good, not perfect, and the artefacts are characteristic. Cymbals and vocal sibilance live in the same frequencies, so a harsh "s" can leave a trace in the instrumental and a hi-hat can bleed into the acapella. Heavy reverb on a voice tends to stay behind in the instrumental, because the tail no longer looks much like a voice. Backing vocals and vocal samples usually go with the lead, whether you wanted them to or not.
The two outputs sum back to the original exactly, because the instrumental is produced by subtracting the vocal rather than being estimated separately. That is a feature — nothing goes missing — but it also means anything the model fails to catch stays audible in the other file.
Making it faster
The page tells you before you start whether it is using your GPU, and estimates how long your file will take. WebGPU is where the speed is: measured on the same machine and the same track, the GPU path runs at about the length of the audio and the CPU path at roughly ten times that. Chrome and Edge on a modern desktop are the safest bet for WebGPU.
The CPU path is multi-threaded where the browser allows it, which is about three times faster than a single thread. The separation itself is identical either way — same model, same numbers out, just a different amount of waiting.
The practical advice is to give it less to do. Trim to the section you need before separating — there is no point processing four minutes to get at one chorus. For anything over the length limit, split it into parts and run them one at a time.
FAQ
Is this real AI separation or just centre cancellation?
Real separation. It runs an MDX-Net model trained to recognise what a voice looks like in a spectrogram, so it works on mono files and on off-centre vocals — both of which defeat centre cancellation. The simple vocal remover is the cancellation approach: instant, where this takes about as long as the song.
Does my song get uploaded?
No. The network is downloaded to your browser and the audio is processed on your machine. The only request is for the model, which is identical for everyone and contains none of your audio. Most free stem splitters upload your track instead.
How long does it take?
Roughly as long as the track when your browser can use the GPU — a four-minute song takes about four minutes. Without WebGPU it runs on the CPU, which measures about ten times slower, so the same song takes closer to forty minutes. The page estimates the time for your specific file before you start.
Why is there a 64 MB download the first time?
That is the trained model. Separation happens on your machine, so the model has to come to you. It is cached afterwards and later songs start straight away.
Do I get four stems like drums and bass?
No — this separates vocals from everything else. Four-stem ONNX models are around 158 MB and will not load in a browser; ONNX Runtime Web aborts partway through. This one is 64 MB and works.
Will it work on my phone?
Possibly, but slowly, and long tracks are likely to run out of memory. Do this one on a desktop.