Vocal remover

Remove vocals from a song in three steps

1. Choose a song (MP3, WAV, M4A, FLAC, OGG or a video file), or load the short original sample.
2. Press Separate vocals. The first time, the model downloads (about 66 MB), then the song is processed in about 10-second pieces with a progress bar and a time estimate. You can cancel at any time.
3. Listen to the vocals and the instrumental, then download either one as MP3 or WAV.

Karaoke maker or acapella extractor: one split, two downloads

Every separation gives you both halves. The instrumental is the karaoke track: the song with the lead voice taken out, ready to sing over or to slow down and practise with. The vocals track is the acapella: the voice on its own, for a remix, a mashup, transcribing lyrics or studying how a singer phrases a line. Because the two masks add up to the original, you can check the split by playing both at once. To sing in a lower or higher key, open the instrumental in the pitch changer; to find which key that is, run it through the song key finder.

How the vocal remover works

The tool uses Spleeter, an open-source source-separation model published by Deezer Research under the MIT license. It turns the song into a spectrogram (a picture of which frequencies sound at each moment), and two small neural networks estimate how much of every point in that picture belongs to the voice and how much to everything else. Those estimates become two masks that are applied to the original sound, so the vocals and the instrumental always add back up to the original song. Older "karaoke" effects subtract the left channel from the right, which only removes a voice panned exactly to the middle and also kills the bass and drums; a trained model does not depend on panning. The model runs with ONNX Runtime Web, on your graphics chip through WebGPU when your browser supports it and on the processor through WebAssembly otherwise.

How fast is it?

In our test on an Apple M2 Mac (headless Chromium), a 3-minute song took 14 seconds with WebAssembly and 5 seconds with WebGPU, after the one-time download. Phones and older computers are slower, often several times slower, so the page measures the first piece of your song and then shows a time estimate. Long songs are processed in overlapping 10-second pieces that are blended together, so memory use stays close to a few copies of the song itself rather than growing with the model.

What quality to expect

Spleeter is fast and small; newer and much larger models separate more cleanly. We tested it on our own synthesized songs (no commercial recordings), where the vocals came out about 18 dB cleaner than the mix, so judge the result on your song. Expect artifacts: cymbals and reverb tails can leak into the vocals, backing vocals and harmonies can end up in either track, and a heavily processed or very quiet voice may be removed only partly. Frequencies above 11 kHz are not analysed by the model; there the tool follows the split the model found just below 11 kHz, which keeps most of that band in the instrumental, so bright breaths and "s" sounds may stay in the karaoke track. Listen to both previews before you download.

Privacy, copyright and offline use

Your song is decoded and separated inside this browser tab. It is never uploaded: AudioClover never receives the audio or its file name. The only downloads are the model files (served from audioclover.com) and the ONNX Runtime engine (from cdn.jsdelivr.net), each checked against a fixed SHA-384 fingerprint before it runs. They are stored in your browser, so later separations, even offline, need no download. Separating a song does not change who owns it: use the results for practice, study and your own projects, and get permission before you publish a track built on someone else’s recording.