Get the acapella out of a song

Lined up, for remixes

The voice comes out exactly as long as the song and starts at the same moment, so it lines up with the original's beat. Export the whole length rather than trimming it, and it will sit on the grid of any music program at the song's tempo. Octoolo keeps the parts as 44.1 kHz FLAC, which is what Demucs, the model it uses, works at.

To keep a short phrase as a sample instead, split the voice at the start and end of the line with Split and delete the rest, then export. Fade the edges by a tenth of a second so it does not click.

How clean is an extracted voice?

The voice is pulled out by a model trained to tell singing from instruments, so it is very good, not perfect. What to listen for:

• Bleed. On loud drum hits or cymbal crashes a faint ghost of them can stay in the voice.
• Effects. Reverb and delay on the original voice come along with it, since they are part of how it was mixed.
• Harmonies. Backing vocals stay in the voice track. They cannot be separated from the lead.

Lowering the voice part's volume a little and adding Reduce background noise can soften a hissy trace. For the cleanest acapellas, start from the best file you have: a WAV or FLAC from the original release separates better than a low-bitrate MP3 or a video soundtrack.

Voices from speech, too

The model treats speech as voice, so the same steps lift a speaker out of a video with music behind it, or an interview with a song playing in the room. Separate the clip's sound, keep Voice only, and the music drops away. Other people talking stay in the voice track, since they are voices too.