Karaoke Studio — AI Stem Separation
Python · FastAPI · Demucs · FFmpeg

A web app that takes any song and splits it with a neural source-separation model into vocals, drums, bass, guitar, piano and the rest, so each part can be muted, practised over or downloaded. Long separations run as background jobs while the browser polls for progress.
Technologies
My role
Sole author. Built the upload and job API, the separation pipeline and the browser player.
Highlights
- Six stems per song from the htdemucs_6s model: vocals, drums, bass, guitar, piano, other
- Optional second pass that separates lead vocals from backing vocals
- Background jobs with status polling, so long separations never block a request
- Bundled FFmpeg, so it runs on a clean Windows machine without extra installs
Screenshots and diagrams
