StemSep

Choose a demo or upload mono/stereo audio. Saved demos open immediately; Run again performs local inference in a browser worker. Recordings use overlapping analysis windows while all stems retain the source timeline.

The current methods compare Spleeter precision variants. They do not represent different model architectures. Exports include raw aligned audio, independent of mixer settings; ZIP exports also preserve source, model and implementation identities.

Weights load on demand from verified cache or a pinned Hugging Face revision. The enabled methods keep uploaded audio on this device. Native inference, commercial services and drum-to-MIDI remain under development.