What they did
The author wanted to reuse the narration from some TikTok videos they had made earlier, but the clips had music under the voice and the original voice recordings were gone. On a Mac, they pasted the video links into Claude Code and asked, in Japanese, for the audio. When they added that the music would likely spoil the result, Claude Code chose Meta's free Demucs separation model on its own. It downloaded the videos, installed Demucs, split the voice track from the music and, on a third request, copied the voice-only files to the Downloads folder. Everything ran locally; nothing was uploaded to a separation service.
How it went
Separation took about 16 seconds for the two videos. The author did not have to look up Demucs or learn how to install it. They note that where music and voice are about equally loud, a little music can still bleed through. They tried only Claude Code, and wrote a more explicit step-by-step prompt for Codex, Grok and Antigravity without testing those.
Worth knowing
Demucs splits audio into vocals, drums, bass and other; the two-stems vocals mode is what keeps speech apart from everything else. It works on speech even though it was built for songs. Only use it on videos you made or have rights to.