Split any song into a clean vocal and instrumental with AI that runs on your own device. Free, no account, and your audio is never uploaded.
The vocal remover splits a finished song back into two tracks: an isolated vocal and a clean instrumental. It runs a neural separation model that was trained to recognise what a singing voice actually sounds like — its harmonics, its formants, the way a sung note moves — rather than relying on the old trick of assuming the vocal sits dead-centre in the stereo image. That's why it still works on mono recordings, off-centre leads, and vocals buried in reverb, and why it leaves your kick and bass intact instead of hollowing out the mix.
Use it for karaoke backing tracks, acapellas to remix or sample, or just to hear a bassline without the singer on top. Results are best on sparse, cleanly produced mixes; very dense or heavily limited masters can leave faint bleed, and starting from a lossless file rather than a low-bitrate MP3 is the single biggest lever you have. Full walkthrough: how to remove vocals from a song, or how AI vocal removal actually works for the mechanism.
Vocal Cut analyzes the full mix and separates it into distinct parts — vocals and instrumental — using a model trained to recognize what a singing voice sounds like, rather than relying on simple tricks like assuming the vocal sits dead-center in the stereo mix. This means it can produce a clean result even on tracks where the classic "center channel" vocal-removal trick would fail, like mono recordings or mixes with an off-center lead vocal. Your device's hardware is used to speed up processing where available, falling back to a slower but still fully local mode if not. Results are generally best on more sparse, cleanly produced mixes; extremely dense or heavily distorted tracks can leave a bit more of the instrumental bleeding into the isolated vocal, and vice versa.
No — audio decoding, stem separation, and file encoding all happen locally on your own device, and your audio bytes and file metadata are never transmitted anywhere. The one network request the tool makes is downloading the AI model itself from a CDN on first use — a download to your device, like any page asset, covered in our Privacy Policy. This makes it a reasonable choice for separating stems from music you're not authorized to redistribute, private session demos, or anything else you'd rather not upload to a third party. To verify it yourself: run one separation, then disconnect your internet — the tool keeps working exactly the same.
Each stem — the isolated vocals and the instrumental — can be downloaded on its own as WAV (lossless, the default) or as Opus in a WebM container (compressed and far smaller, ideal for phones and sharing). The estimated file size is shown next to each option and you choose the format per download, so you could grab a lossless vocal for further editing and a small Opus instrumental to share. WAV is the safe pick if you'll keep editing the stem; Opus is best when you just need a compact file.
The tool gives you two stems: a clean vocal track and a single instrumental that combines drums, bass, and everything else. The model does distinguish those parts internally, but they're summed into one instrumental for a simpler, faster result on your own device. For making a karaoke backing track or pulling an acapella, two stems is exactly what you need; full multitrack separation into individual drum and bass stems is a much heavier job better suited to dedicated desktop software.