01 / Drop
Choose a song
Any song or a music video, up to 15 minutes.
Free tool
Split a song into four stems for a remix, a cover, practice or sampling. An AI model trained to separate music runs in this tab, so your song is never uploaded. Each stem downloads as MP3 or WAV, or all four in one ZIP. No sign-up, no watermark.
This runs in your browser. Your files are never uploaded.
How it works
01 / Drop
Any song or a music video, up to 15 minutes.
02 / Split
The song is pulled apart into four tracks, about 8 seconds at a time.
03 / Save
Listen to each one, then save the ones you need or all four in a ZIP.
Questions
Four: vocals, drums, bass, and everything else together (guitars, keys, synths, strings). Played together, they sound like the full song again. Piano or guitar on their own are not split out.
No. The AI model is downloaded to your browser, from Hugging Face, the open model library, and runs there, so the file never leaves your computer. No account and no limit on the number of songs.
HTDemucs, the music separation model by Meta AI researchers, released under the MIT licence, in a browser version by the demucs-web project. It is the same model as our vocal remover, which mixes the drums, bass and other stems back into one instrumental.
Good, but not perfect: a trace of one part can stay in another, for example cymbals in the vocals or a bass note in the drums. We have not measured it on real songs, so listen to each stem before you use it.
It depends on your computer. With WebGPU, which uses the graphics card, a 20 second song took about 5 seconds in our test; without it, the processor is much slower. The first time, the model downloads, 181 MB. Songs up to 15 minutes.
For practice, study or your own songs, yes. To publish a remix or a sample of someone else's song you need the rights holder's permission.