How to remove vocals from a song
- Drop your song. Drop it into the box, or press Try a sample. MP3, WAV, FLAC, M4A and videos all work.
- Press Split vocals. The first time, the model downloads (about 20 MB); after that it starts at once. The bar shows how long is left.
- Listen and compare. Solo or mute each track, and press B to flip between your original and the separated version.
- Download. Save the instrumental (karaoke) or the vocals (acapella) as lossless WAV or a smaller MP3.
What makes it different
Most vocal removers send your song to a server. This one separates it in your browser, so nothing leaves your device.
No account, no watermark, no daily cap, no low-quality preview: full 44.1 kHz stereo WAV.
SCNet, a 2024 model in the same quality class as Demucs, not the old "cancel the centre channel" trick that wrecks the instruments.
Every track plays in sync. Solo, mute and switch to the original without a gap to hear exactly what was removed.
Stem splitter: what the stems are good for
Take the acapella into a new beat, or the instrumental under a new topline.
Mute the part you sing and sing along, or solo the voice to learn the melody by ear.
The instrumental is a backing track for singing, rehearsing or a cover video.
The vocals on their own are much easier to write down, by ear or with speech to text.
The instrumental is made as your song minus the separated vocals, so the two always add back up to the original, sample for sample. Any separation error shows up as a little bleed of one into the other, never as a missing piece.
Four stems with Pro
The free tool gives two stems: the vocals and everything else. Pro will split the instrumental further into drums, bass and other, from the same model run, on your device like everything else here. It is not on sale yet: there is nothing to pay for today, and vocals plus instrumental stay free with no limits.
Quality: Standard or Best
The model hears a song in overlapping 11-second sections. Standard hears each moment twice; Best hears it four times and blends the results. Best takes about twice as long for a very small gain (on our test song, 0.03 dB), so Standard is the right choice for almost everything.
How long it takes
Measured on a desktop (Intel Core i7-14700KF, GeForce RTX 5060 Ti) in Chrome, Standard quality. Laptops and phones are slower; the progress bar always shows the real time left.
| Engine | Per 11 s section | 4-minute song |
|---|---|---|
| Graphics card (WebGPU) | 3.5 s | about 3 min |
| CPU, 8 threads | 4.4 s | about 3.5 min |
| CPU, 4 threads | 10 s | about 7.5 min |
More audio and video tools
- Noise remover: take fans, traffic and hum out of a voice recording, or clean up the acapella.
- Auto captions and video to text: lyrics or speech to SRT, VTT or plain text, in your browser.
- Video compressor: make a cover or karaoke video small enough for Discord, WhatsApp or email.
- Text to speech: a natural voice-over to lay over the instrumental, made on your device.
Privacy
Your song is decoded and separated by your browser on this page. It is not uploaded, stored or seen by anyone. The page downloads the model and the ONNX Runtime engine (about 20 MB compressed) from this site once and keeps them in your browser's cache.
Questions
Is this vocal remover really free?
Yes. Splitting a song into vocals and instrumental is free, with no sign-up, no watermark and no daily limit, for songs up to 15 minutes. Downloads are full-quality WAV. Splitting into four stems (vocals, drums, bass and other) will be part of Pro.
Is my song uploaded anywhere?
No. The separation model is downloaded to your browser once and runs on your own computer or phone. Your song never leaves your device, and once the model is cached the tool even works offline.
How good is the separation?
It uses SCNet, a 2024 research model that scores about 9.9 dB SDR on vocals on the MUSDB18 benchmark, in the same class as Demucs. On a real Creative Commons song we checked against the artist's own instrumental, the instrumental came out at 16 dB SDR and the vocals at 9 dB. Expect faint reverb tails or backing vocals now and then; very old mono recordings and heavily distorted mixes are the hardest.
Is this a stem splitter?
Yes, for two stems: vocals and instrumental, free. Four stems (vocals, drums, bass and other) come from the same model and are coming with Pro.
Why is it slow on my computer?
The tool is fastest when your browser can use the graphics card through WebGPU (Chrome or Edge on desktop, Chrome on Android, Safari 26). Without it, the same model runs on the CPU, which is several times slower but still works. The progress bar shows the time left, and the help box under the tool explains how to turn graphics acceleration on.
What files can I use?
MP3, WAV, FLAC, M4A, AAC, OGG and Opus, and videos (MP4, WebM, MOV) with an audio track. You get the vocals and the instrumental as 44.1 kHz stereo WAV or MP3.
Can I make a karaoke track or an acapella?
Yes. The instrumental is a karaoke or backing track: everything except the singing. The vocals track is an acapella you can use for remixes, cover practice or transcription.
WAV or MP3?
WAV is lossless, so nothing is lost on the way out, and every audio app and editor opens it. MP3 (about 320 kbps) is a fifth of the size and hard to tell apart; use it for sharing and practice.
Does it work on a phone?
Yes, on recent Android phones in Chrome and on iPhone and iPad with Safari 26. Long songs need a lot of memory, so on a phone we suggest songs under about six minutes and closing other tabs first.
Which AI model does it use?
SCNet (Tong et al., ICASSP 2024), released under the MIT licence by its authors, in an int8 ONNX build of 12.9 MB run by ONNX Runtime Web. Credits and licences are listed at the bottom of this page.
Can I use the separated tracks commercially?
The tool puts no restrictions on what you make, but the song itself still belongs to its rights holders. Removing the vocals does not change the copyright: get permission before you publish or sell anything made from someone else's song.
Credits and licences
- SCNet by Weinan Tong, Jiaxu Zhu, Jun Chen, Shiyin Kang, Tao Jiang, Yang Li, Zhiyong Wu and Helen Meng (ICASSP 2024), starrytong/SCNet, MIT licence (licence text; all third-party notices: NOTICE). The authors confirmed the released weights are MIT too.
- Checkpoint distributed by Music-Source-Separation-Training (Roman Solovyev, MIT); ONNX export by audiojs (MIT).
- ONNX Runtime Web (MIT) runs the model. Audio is read and written with Mediabunny (MPL-2.0).
- MP3 files are made with the LAME MP3 encoder 3.100 (LGPL-2.0-or-later, unmodified, as WebAssembly), loaded as a separate file only when you download an MP3: its source, licence and how to replace it (LGPL text).
- Sample: "Start Again" by Alex Beroza feat. Snowflake and Subliminal, CC BY 3.0, a 10-second excerpt with fades.