Remove the vocals from a song

What runs on your computer

Online vocal removers upload your song to a server, put you in a queue and often ask for credits for anything long or in high quality. Octoolo does the same job on your own PC with Demucs, an open-source separation model from Meta's AI research team (the Hybrid Transformer version). The song never leaves the computer.

The model is a one-time download of 91 MB, plus the 15 MB AI engine if no other AI tool in Octoolo has fetched it yet. Both are checked against their fingerprints and kept, and you can remove them again under Settings. After that it works offline. The Audio Editor also needs FFmpeg, a one-time 77 MB download, for reading and writing sound files.

How long it takes

Octoolo uses the graphics card when it can, through DirectML, and the processor otherwise. When a card is not clearly fast, it times both once and keeps to the quicker one. On our test PC a five-minute song took about 14 seconds on an NVIDIA RTX 4060 and about two minutes on its 20-thread processor alone. A laptop without a graphics card takes longer, so expect a few minutes for a song there.

The window shows how far along it is and roughly how long is left, and Stop ends it at any point. Long recordings work too: the time grows with the length, while the memory it uses stays about the same. Separating the same song again is instant, because the parts are kept.

What to expect from the result

On a typical studio recording the voice comes away cleanly. A few things decide how clean:

• Backing vocals and harmonies count as voice. Turning the voice off removes the choir as well as the lead singer.
• Reverb and echo on a voice can leave a faint trace in the music, most noticeable in quiet passages.
• Instruments that sound like a voice, such as a lead synth or a saxophone, can lose a little to the voice track.
• Live and old recordings with the voice in every microphone separate less cleanly than studio tracks.

The music track is the song minus the voice, so the two together are always exactly the original. If something sounds wrong, switching the voice back on puts the song back as it was.

Two parts or four

Voice and music is what most people want: a karaoke or instrumental track with the voice off, or an acapella with only the voice. Voice, drums, bass, other splits the music further, for musicians who want to practice along without the bass, or to remix: see stem splitter. Both come from the same run of the model, so choosing four parts takes no longer.