Ultimate Vocal Remover online alternative: the same separation, in your browser

Ultimate Vocal Remover, usually written UVR or UVR5, is an open-source desktop application for Windows, macOS and Linux. It is a front end for a shelf of source-separation models: MDX-Net, its own VR Architecture networks, Meta's Demucs, and more recently the RoFormer family. You pick a model, or an ensemble of several, and it runs on your machine, quickly if you have a recent NVIDIA graphics card and slowly if you do not. It is free, it is good, and a lot of the vocal-removal tutorials on the web are really UVR tutorials.

What UVR does not have is a website. The developers publish a program you download and install, not a page you open. Every site that advertises "Ultimate Vocal Remover online" is a third party that took some of the same public models, put them on a server, and added an upload form, a queue and a quota. Some of them are fine. None of them is UVR, and all of them have your song for as long as they keep it.

The separation itself is the same kind of work. SplitVocals runs Demucs v4, the hybrid transformer model that UVR offers under its Demucs tab, converted to run through WebGPU on your graphics card or through WebAssembly on the CPU when WebGPU is not available. The four stems it returns (vocals, drums, bass, other) are the four UVR's Demucs option returns. The model downloads once, about 170 MB, and lives in your browser cache; the song you drop in is decoded and separated on your device and is never sent anywhere.

What is different is everything around the model. UVR gives you a menu of dozens of networks, ensemble modes that run several and average them, and knobs for segment size, overlap and post-processing. This page gives you one browser model with sensible settings and a button. For the tracks where one browser model is not enough, the HD Pack and Pro plans run a larger vocal model of the RoFormer family plus the six-stem Demucs on a server GPU, and return WAV or FLAC; that is the one case where a file leaves your device, and it is deleted about an hour later.

If you have a modern NVIDIA card and you separate audio every day, install UVR. Ensembles of its best vocal models still edge out a single Demucs pass on the hardest vocals, it has specialised networks for jobs a general separator does not do (de-reverb, de-echo, splitting lead from backing vocals), and it will chew through a folder of two hundred files overnight without a browser tab open. It is also the tool to learn if you want to understand what the models are doing, because it exposes all of it.

Most people who search for UVR online are not in that position. They have one song to do tonight, a laptop they cannot install software on, a Mac with no NVIDIA card, or a Chromebook. For them the browser is the practical route: no multi-gigabyte install, no model files to hunt down, no Python errors, and no server holding the audio. Apple silicon Macs and any recent Windows laptop with a GPU run the WebGPU path in one to three minutes per song; older machines take the CPU path and simply take longer.

The privacy point is worth stating plainly. The online "UVR" services and the desktop program sit at two ends: one uploads everything, the other uploads nothing. The browser separator is on the desktop program's end of that line, with the convenience of the other.

Drop a file on the box above: MP3, WAV, FLAC, M4A, OGG, AAC or WebM, up to 50 MB or 10 minutes. The first run fetches the model; later runs start at once, even offline. Pick the stems you want as they finish, or download all four. Use the studio version of a song rather than a live take, and a lossless file when you have one; both matter more than which tool you use.

If the vocal comes back with a halo of reverb or the instrumental keeps a ghost of the voice, the song is one of the hard ones. That is when the HD tier earns its price, and it is also the point where a UVR ensemble on a good GPU would do better than a single model. Neither can undo a live recording or a 128 kbps MP3.

On the standard research benchmark, Demucs v4 sits around 9 dB of signal-to-distortion ratio, comfortably above the Spleeter generation and close to the strongest single models UVR ships. UVR's top ensembles and the newest RoFormer vocal models score higher on vocals specifically, which is why the HD tier here uses a RoFormer model rather than a bigger Demucs. For a karaoke track, an acapella to build a mashup on, or drums to sample, the browser result is usable as it is; for a commercial release, a producer will want the HD pass or a desktop ensemble, and probably some clean-up either way.