AI Vocal Remover
A vocal remover that splits a song into two stereo tracks, the voice on one and the instrumental on the other, using Deezer's Spleeter model.
The song
Song
Drop the song here, or choose a file
MP3, M4A, WAV, FLAC, Ogg — anything the browser decodes
—
—
What comes back, and what it costs
Where the time goes
Add a song and its passes are counted here, 11.9 s of the song each.
Vocals and instrumental
Two stems come back: the voice on its own, and the band without it. Add a song and the cost above gets its real numbers.
How to use
- Pick the song. The Song field takes anything your browser can decode, so MP3, M4A, WAV, FLAC and OGG are all fine, and the plan below fills in from the decoded audio.
- Choose the output format first. Both stems are written in it as the split's last step.
- Press Split vocals. The first run fetches the model, and the badge above the player says how much that is and whether it's already here. Then the bar counts the passes, and Cancel is right there if you grabbed the wrong track.
- Listen with the Take toggle. Original, Vocals and Instrumental play from the same moment on one player.
- Download a single stem or take both in one zip. Change the format afterwards and both stems are written again before the downloads come back.
How the split works
Feed it a song and you get two files back: the voice on its own, and the band without the voice. The engine doing that is Spleeter, the separation model Deezer open-sourced, running on this device: on the graphics card through WebGPU where the browser offers it, on the processor where it doesn't. Both stems come back in stereo at 44.1 kHz, and one player switches between the original and the two stems at the same moment, so you can judge the split before you save anything.
Be realistic about the grade. Spleeter dates from 2019, and it is the one separation model whose training data is cleared for commercial use, so what you get is karaoke and practice quality rather than release quality. Downloads are M4A, WAV, Opus or FLAC, chosen before the split and written once at its end, one stem at a time or both together in a zip.
A four-minute pop track, start to finish
Drop in a 4:07 MP3. Before you press anything, the plan under the format has already counted it from the decoded audio: 10,634 frames, split into 21 passes of 11.9 s. Press Split vocals. The first time, the bar counts the model files in megabytes, then reads Separating — pass 1 of 21 and fills one pass at a time. With WebGPU the whole thing is over in seconds; on the processor it is slower but still finishes in a fraction of the song's own length, and the badge says which one ran.
The player lands on the Vocals take at the moment you were listening to, the lead and the harmonies with a ghost of the snare behind them. Press Instrumental and the same second plays without the singing, which is your karaoke track. Leave the format on M4A, press Download both (.zip), and the archive is named after the file you started with, holding one file ending in -vocals.m4a and one ending in -instrumental.m4a.
Good to know
- Faint bleed is normal. A dense mix leaves a trace of the lead in the instrumental, and a busy chorus leaves a trace of the band in the vocals. There's no manual touch-up afterwards.
- Two stems and no more. Drums, bass and guitar all land together in the instrumental, so you can't get a per-instrument breakdown out of this.
- Songs longer than 20:00 are held before the split. 40:00 was measured to finish on a laptop and the page holds at half that, so cut a longer recording into pieces in the Audio Cutter first.
Frequently asked questions
Will this give me a karaoke track good enough for a party?
On most songs the lead drops far enough that a room singing over it will never notice what's left. Quiet, close-miked ballads separate best. A wall-of-sound chorus with doubled vocals and heavy reverb is where you'll hear traces.
How does it actually pull the voice out of the mix?
The song becomes a spectrogram, and two small neural networks look at that picture and estimate how much of the energy in each time-frequency cell belongs to a voice. Those estimates become soft masks, one per stem, and each masked spectrogram is turned back into audio.
Can I use the vocals stem in a remix?
It's good enough for sampling, sketching an idea, or a bootleg edit. Two things show up if you push it. Both stems stop at 11,025 Hz: Spleeter's two-stem model only looks at the bottom 1,024 bins of a 4,096-point transform and leaves everything above them silent, so cymbals and air lose their sparkle and the vocal sounds slightly veiled next to a real acapella. Busy passages also carry some of the band along with the voice.
Why can't I download an MP3?
Every MP3 encoder I could ship carries a licence this site's rules don't allow. M4A is the drop-in replacement and plays wherever an MP3 would. WAV and FLAC are there for stems headed into a DAW, and Opus when you want the smallest file.
How long does one split take?
Less time than the song lasts, on anything modern. A three-minute track goes through in a few seconds when your browser exposes WebGPU; the processor path is slower and still comes in well under the track's own duration. The plan counts the passes for your song before you press, and once a split has run on the processor the badge says how long a pass took on your device.
What happens the very first time I use it?
Two model files arrive, one for vocals and one for the accompaniment, 18.8 MB apiece, plus the runtime that runs them; the badge above the player adds it all up for your device. That's the only network traffic the tool makes. Once they're cached, later splits start straight away and the page keeps working with the connection off.
Can I run this on a phone?
It runs, but a long track asks a lot of a phone. The audio is held in memory as 44.1 kHz stereo, and the plan under the format shows how much spectrum your song holds while the passes run. Past 20:00 the page holds the split and points you to the Audio Cutter; that limit comes from a laptop, and a phone with less memory may stop sooner.
Missing something in AI Vocal Remover — Split Vocals & Instrumental? Suggest a feature →
Tell someone who needs this.