Drop a song and play it without the vocals seconds later. Keep only the voice, or split out the drums, bass, and instruments to practice, remix, or sing along. The AI runs in your browser, so nothing is uploaded.
This is local AI. The same approach runs private assistants, search, and document processing on your own hardware, with no data sent to an AI provider.
Local AI for your businessHow it works
Any MP3, WAV, M4A, FLAC, or video, split into 4, 6, 8, or 10 stems. The first run downloads the AI model once (94 MB for 4 stems); after that it starts instantly.
Stems are separated from the start of the song onward, and you can listen within seconds while the rest is still splitting.
No vocals, vocals only, no drums, or any mix of your own, as WAV. Every stem also comes as MIDI, with the song's BPM, key, and chords, for FL Studio, Ableton, or your keyboard.
Why use it
One click for the instrumental, one for the voice alone. Great for singing along, covers, and remixes.
Mute the drums, bass, or guitar and play along with the band, or solo a part to learn it note by note.
Drag across the song to loop a riff or a verse, and hear it again and again with your part taken out.
The melody, bass line, piano, guitar, and drums come out as MIDI notes you can edit, re-voice with any instrument, or open as sheet music in MuseScore.
One General MIDI file for your keyboard or digital piano, melody on channel 1, so you can mute it and play the tune yourself over the band.
The chords scroll by as the song plays, with the key and tempo shown. Download a chord chart, bar by bar, to play along in any style.
Every stem as a 16-bit WAV and MIDI, lined up on the beat grid at the song's tempo, in one ZIP for FL Studio, Ableton, Logic, or GarageBand.
Most vocal removers give you 30 seconds free, then charge. This splits the whole song, as often as you like.
Unreleased demos and client work stay on your computer. There's no upload, so there's no queue and no copy on a server.
The Vocal Remover runs HTDemucs, Meta's open source-separation model, on your device: on its graphics chip when the browser allows it, otherwise on its processor. For 8 and 10 stems it adds two more open models: UVR's karaoke model splits the vocals into lead and backing, and drumsep splits the drums into kick, snare, cymbals, and toms. The MIDI comes from Spotify's open Basic Pitch model, and the tempo, key, and chords from signal analysis, also on your device. Your songs, stems, and MIDI are never uploaded, so there's nothing to store, share, or delete later. The models themselves download once, from this site. We count anonymous usage (file name, type, and size, and browser) to keep the tool improving.
Read exactly what we recordFAQ
Drop the song above and press play: in a few seconds you'll hear it with the vocals taken out. Pick “No vocals” and download the instrumental as a WAV. The whole song is free, with no sign-up.
Yes. The separation runs on your own device, not on our servers, so there's nothing to ration: full songs up to 10 minutes, as many as you like.
No. The AI model runs in your browser, on your own device, so the music never leaves it. Only the AI model itself is downloaded, once, from this site.
Pick “Vocals only”, then “Download this mix”. You can also download the vocal stem on its own from its row.
You choose: 4 stems (vocals, drums, bass, and everything else), 6 (adds guitar and piano), 8 (the vocals split into lead and backing, and the drums into kick, snare, cymbals, and toms), or 10 (all of it: lead and backing vocals, kick, snare, cymbals, toms, bass, guitar, piano, and everything else). Switching from 4 to 8, or 6 to 10, keeps what's already separated and only splits further.
Yes. Pick 8 or 10 stems, then “No lead vocal”: the harmonies and choirs stay in, which is what most karaoke tracks sound like.
It uses HTDemucs v4, the model behind many paid stem splitters. Vocals and drums come out very clean; heavy effects, live recordings, and unusual instruments can leave a little bleed between stems.
Your browser downloads the model the first time and keeps it, so the next song starts right away: 94 MB for 4 stems, 56 MB for 6, and about 315 MB for 8. It downloads in small parts, so a dropped connection picks up where it stopped. The first seconds of a song also include starting the AI.
Yes. It runs on your graphics chip where the browser allows it (Chrome and Edge on Windows and Mac, recent Android and iPhone browsers), and on your processor everywhere else, which is slower but always works. A computer with a graphics card is fastest.
MP3, WAV, M4A, AAC, FLAC, OGG, AIFF, and videos (MP4, MOV, WebM): anything your browser can play. Stems download as WAV and MIDI, and the chords as a text chart.
Yes. Once a song is split, every stem is turned into MIDI on its own: the vocal melody, the bass line, the piano, guitar, and other instruments as notes, and the drums as a General MIDI drum track. Download one stem's MIDI from its row, or all of them at once. Separating first is what makes it work: each instrument is transcribed alone, not the whole mix at once.
Pick “Download project”: a ZIP with every stem as WAV and MIDI, the whole band in one MIDI file, the chords, and a read-me with the song's tempo. Set your project to that tempo and drop the files in at bar 1. A little silence is added at the start so the first downbeat lands on a bar line, so the grid matches the music.
Yes. “Download MIDI” under Keyboard gives one General MIDI file that keyboards play from a USB stick, with the melody on channel 1 (the right-hand part on most keyboards), so you can mute it and play the melody yourself. The chord chart gives the chords bar by bar to play along with any of your keyboard's styles.
By listening to the separated stems: the tempo and beats from the drums, the key from everything but the drums, and the chords from the harmony without the voice, checked against the bass. It's usually right for pop, rock, and dance music; songs that change key or speed up and slow down are harder, so trust your ears on the tricky bars.
For practice, study, and personal karaoke, go ahead. To publish a remix or cover, you still need permission from whoever owns the song, whatever tool made the stems.
More free tools
Passport, visa, and upload photos
Photo Resizer
Pick where the photo is going (passport, visa, or a size limit) and get a file that passes, first time.
OpenStrip hidden data from files
Metadata Remover
Remove GPS location, camera details, and author names from photos, PDFs, and documents.
OpenSee what a file reveals
Metadata Viewer
See where a photo was taken, which device made it, who wrote a document, and whether it says it's AI.
OpenWant tools like this in your own product?
Eligapris builds fast, private web apps and on-device AI. The same engineering behind Vocal Remover, for your team.