This vocal remover splits a song into two tracks — the vocals on their own and the instrumental without them — using Spleeter, the open-source source-separation model from Deezer Research. Use the instrumental as a karaoke or practice backing track, or the vocal track as an acapella for a remix, a cover or transcribing lyrics. It reads MP3, WAV, M4A, OGG and FLAC files and the soundtrack of videos, and you can switch between Original, Vocals and Instrumental at the same point in the song before downloading either one as WAV or MP3.
Everything runs in your browser. The first separation downloads the model (two networks, 39 MB in total) and the 24 MB AI engine shared with our other AI tools, after showing you the size; both are then cached by the browser; after that, separating a 4-minute song took 7 seconds on a recent laptop when the browser could use the graphics card (WebGPU) and 35 seconds on the processor alone; older computers and phones take longer. Spleeter is a 2019 model: it separates most pop, rock and hip-hop cleanly, but you will hear some leftover voice in the instrumental on dense or reverb-heavy mixes. There is also an instant classic karaoke mode that subtracts one stereo channel from the other; it is not AI and sounds rougher, but needs no download. To trim a stem afterwards, use the audio cutter; to pull the sound out of a video first, video to MP3.
How to use
- Drop a song or a video onto the box, or click to choose one.
- Keep AI separation selected (or choose Classic karaoke for an instant, rougher result).
- Click Separate. The first time, the model downloads; then a progress bar shows each part of the song being processed.
- Press Play and switch between Original, Vocals and Instrumental to check the result.
- Choose WAV or MP3 and click Download vocals or Download instrumental.
Frequently asked questions
How do I remove vocals from a song for free?
Open the song on this page, keep AI separation selected and click Separate. The Spleeter model splits it into vocals and instrumental in your browser; download the instrumental as WAV or MP3. There is no account, no watermark and no monthly limit; songs can be up to 10 minutes long.
Is this really AI, and which model is it?
The default method is Spleeter 2-stem, a U-Net neural network published by Deezer Research under the MIT licence, trained on studio multitrack recordings to predict which parts of the spectrum belong to the voice. It runs here with onnxruntime-web. The Classic karaoke option is not AI: it subtracts the right channel from the left, which removes anything mixed exactly in the centre.
Why can I still hear some vocals in the instrumental?
No separator is perfect, and Spleeter (2019) is lighter than the newest models such as HTDemucs or MDX-Net, which are several times larger (HTDemucs is about 180 MB as one file) and too heavy to download and run in a web page comfortably. Expect faint vocal traces on reverb-heavy, choir or distorted-guitar tracks, and breathy or 'phasey' artefacts in the vocal track. Results are usually good enough for karaoke and practice.
Can it split drums, bass and other instruments too?
Not at the moment. This page separates two stems, vocals and instrumental. Four-stem models (vocals, drums, bass, other) are at least twice the download, so they are not offered in the browser here.
Is my song uploaded to a server?
No. The model files are downloaded from this site to your browser and the song is processed on your device; nothing you open is sent anywhere. Unusual file types the browser cannot read on its own are decoded with ffmpeg, which is also downloaded (from the jsDelivr CDN) and runs locally.
Can I use the separated tracks however I like?
Separating a song does not change who owns it. Practising, karaoke at home and studying an arrangement are generally fine; publishing or selling a remix, cover or backing track made from someone else's recording needs permission from the rights holders.
AI separation vs the classic karaoke trick
The classic method relies on how songs are mixed: the lead vocal is usually placed dead centre, so it is identical in the left and right channels, and subtracting one channel from the other cancels it. So does everything else in the centre — bass, kick drum, snare, often the lead instrument — which is why this tool adds the bass below 120 Hz back. What survives is whatever was panned to the sides, plus the vocal's stereo reverb as a ghostly echo. It is instant and needs no download, but it only works on stereo files and gives no vocal track.
Spleeter works differently. It turns the song into a spectrogram (how loud each frequency is, every 23 milliseconds), and two neural networks, trained on studio multitracks, estimate how much of each point belongs to the voice and how much to the accompaniment. Those estimates become soft masks that are applied to the original sound, so the instrumental keeps the bass, drums and centre instruments, and you also get the vocals on their own.
How the browser version is built
The model sees about 12 seconds at a time, so the song is processed in overlapping parts; where two parts overlap, their masks are cross-faded so there is no audible seam. Spleeter's networks only look at frequencies up to 11 kHz; above that, each moment gets the average of its masks, so cymbals and sibilance are shared between the stems instead of being cut off. The instrumental is the original minus the vocals, which is exact because the two masks always add up to the whole.
The networks run with onnxruntime-web in a background worker: on the graphics card through WebGPU in browsers that support it (such as current Chrome and Edge), otherwise on the processor through WebAssembly. Both give the same result; only the speed differs.
Tips for a cleaner split
• Start from the best source you have: a WAV, FLAC or 320 kbps MP3 separates more cleanly than a low-bitrate file, whose compression artefacts confuse the model.
• For karaoke, a little leftover voice in the instrumental is masked once you sing over it; check the quietest passages, where traces are most audible.
• For an acapella, the vocal track may carry some drum hits or guitar; trim it, or lower those parts in an audio editor.
• Songs longer than 10 minutes can be cut into parts with the audio cutter first.
