# Vocal Remover > Remove vocals from a song or pull out the acapella with an AI model on your graphics card. Preview 20 seconds, then save both tracks as WAV. No upload. - URL: https://bytescope.dev/en/vocal-remover/ - Locale: en · Also: [ja](https://bytescope.dev/ja/vocal-remover.md) · [zh-Hant](https://bytescope.dev/zh-Hant/vocal-remover.md) - Runs: Your files never leave your browser — all processing is local. - Requires: A desktop browser with WebGPU (tested in Chrome) on a real graphics card with more than 4 GB of video memory (the model peaks around 3.5 GB). It does not run on the CPU: without WebGPU, or with a software renderer such as SwiftShader, the page explains why and does not start. The first run downloads a 175 MB model and a 26 MB engine once; phones are not supported. - Updated: 2026-09-08 ## What it does A vocal remover splits a finished stereo mix back into parts. Drop an MP3, M4A, WAV, FLAC or OGG file — or a video, whose soundtrack is used — and the page returns two tracks: the voice alone (an acapella) and everything else (an instrumental you can sing over). Singers use the instrumental for karaoke and cover practice, producers use the acapella as the start of a remix, and musicians use either to hear what is really going on in a dense arrangement. Everything happens inside this browser tab: the file is decoded here, the model runs on your own graphics card, and nothing is uploaded. ### How the separation works The model is HTDemucs, the Hybrid Transformer Demucs published by Défossez and colleagues at Meta in 2022 and released under the MIT licence. It reads the song both as a waveform and as a spectrogram, and a transformer lets the two views inform each other. It really separates four sources — drums, bass, vocals and everything else — and this page adds drums, bass and the rest back together to make the instrumental. The song is resampled to 44.1 kHz stereo and cut into 7.8-second windows that overlap by a quarter; each window goes through the model on the GPU with ONNX Runtime Web, and the windows are cross-faded back together so there are no seams. The spectrogram maths in and out of the model runs in a Web Worker, so the page stays responsive during a long song. ### Preview first, then the whole song The first button separates 20 seconds from the middle of the song, which usually lands on a verse or a chorus, so you can hear how well this particular song separates before spending a minute or more on all of it. When it sounds right, separate the whole song: the progress line shows how many seconds are done, the speed as a multiple of real time and the time left. On a mid-range card such as a GTX 1060, a three-minute song takes about 45 seconds. Both tracks are saved as 16-bit stereo WAV files at 44.1 kHz, exactly as long as the original. ### What to expect Separation is not perfect. Expect a little instrument bleed in the vocal track, and a trace of backing vocals or reverb left in the instrumental — most of all with stacked harmonies, choirs, and vocals soaked in effects such as heavy reverb, delay, distortion or auto-tune, where the model cannot tell which part of the sound belongs to the voice. Sparse pop and rock mixes separate best. Speech over music works too, but the model was trained on songs, not podcasts. ### What your computer needs The first run downloads the model once, about 175 MB, plus about 26 MB for the inference engine (ONNX Runtime Web, from the jsDelivr CDN); both are then kept in this browser's cache. The model only runs on a real graphics card through WebGPU and needs about 3.5 GB of video memory at its peak, so a card with 4 GB or less may fail. There is no CPU fallback: on the processor the same work took several times the length of the song and nearly 3 GB of memory. If your browser has no WebGPU, or WebGPU is running on a software renderer because graphics acceleration is turned off, the page says so instead of starting, and links to My Computer Specs so you can check. It is made for desktop computers; phones and tablets are not supported. ### Copyright The separated tracks are still the original song, and still under its copyright. Use them for personal practice, study or private karaoke; do not publish them, share them or upload them to streaming or video sites without permission from whoever holds the rights. ## Limits Does not upload songs, does not separate perfectly (some instruments bleed into the vocal track and some backing vocals or reverb into the instrumental, especially with harmonies or heavy vocal effects), does not split drums, bass or other instruments yet, does not write MP3 (WAV only), and does not clear the copyright of the original song. ## FAQ ### Is my song uploaded? No. The file is decoded in this tab, and the separation runs on your own graphics card through WebGPU. The only downloads are the model and the inference engine, the first time; your audio never leaves the browser and nothing about it is recorded. ### Why does it need a graphics card? HTDemucs is a large network — the model file alone is about 175 MB — and every 7.8 seconds of music takes billions of calculations. On a GTX 1060 through WebGPU a three-minute song takes about 45 seconds. On the processor the same work was several times slower than the song itself and used nearly 3 GB of memory, so instead of a CPU mode that would freeze the tab, the page tells you what is missing and links to My Computer Specs to check your browser and graphics card. ### Why won't it run on my phone? At its peak the model needs about 3.5 GB of video memory on top of the browser itself, more than phones and tablets give a web page, and the first run downloads about 200 MB. The page is built for desktop and laptop computers with a real graphics card; on a phone it explains this rather than starting a job that would crash or take far too long. ### How big is the download, and where is it kept? About 175 MB for the model and about 26 MB for the inference engine, once. The model is checked against its SHA-256 fingerprint and stored in this browser's cache, so the next song starts without downloading anything. Clearing this site's data in your browser settings removes it. ### Why can I still hear a little of the voice in the instrumental? Because a finished mix does not keep the parts apart — reverb, harmonies and effects smear the voice across the whole sound, and the model has to guess which part of each sound belongs to the singer. Stacked backing vocals, choirs, and vocals with heavy reverb, delay or auto-tune leave the most behind; a clean lead vocal over a sparse arrangement separates best. Use the 20-second preview to judge a song before separating all of it. ### Can it split drums, bass and the other instruments too? Not yet. The model already separates drums, bass, vocals and everything else internally, but this version gives you two tracks: vocals, and an instrumental that adds the other three back together. Separate drum, bass and other tracks with a mixer are planned for a later version. ### Can I use the separated tracks in my own video or release? Only with the rights holder's permission. Separating a song does not change who owns it: the acapella and the instrumental are still parts of the original recording. They are fine for personal practice, study and private karaoke; do not publish, distribute or upload them to streaming or video sites without a licence. ## Related tools - [BPM & Key Finder](https://bytescope.dev/en/bpm-key-finder.md) — Find the tempo and key of the song before you sing or play over it. - [Audio to MIDI](https://bytescope.dev/en/audio-to-midi.md) — Turn the separated vocal melody into a MIDI file. - [Extract Audio from Video](https://bytescope.dev/en/extract-audio.md) — Pull the soundtrack out of a video as an audio file.