Make an acapella from any song you have. The separation model splits the vocals from the instruments on your own device, and you download the vocals as MP3 or WAV, with the instrumental too if you want both halves.
How it works
- Drop a song: MP3, WAV, M4A, FLAC, or a video whose soundtrack you want.
- Pick what to keep. The model downloads once (about 90 MB) and is cached.
- Listen to each part in the page, then download it as MP3 or WAV.
Questions
- Does my file get uploaded?
- No. The models and all processing run inside your browser on your own device. You can verify this yourself: open your browser’s DevTools, watch the Network tab, and process a file: there are no uploads. After the first visit most tools also work fully offline.
- How clean is the separation?
- Very good, but not studio-clean. Demucs is one of the strongest open models for this, and most songs come apart well, but faint traces of the other parts can remain, especially reverb on the vocals and cymbals that share their frequencies. It works from the mixed song alone, so it cannot recover what the mix covered up.
- How long does it take?
- Measured on a laptop: with a supported GPU, about half the song’s length (a three-minute song in roughly a minute and a half). Without one it runs on the processor at about 1.8× the song’s length, so around five minutes for the same song. The page tells you which path your device is on before you start.
- Which model is it, and can I use the results?
- HTDemucs v4 from Meta AI Research, released under the MIT licence, exported from Meta’s own checkpoint. The licence covers the model; the song is still whoever owns it, so a stem of a commercial track is for your own practice, not for release.