maulay/music-splitter-mixit

Mixture-only (MixIT) music source separator — a torchaudio ConvTasNet trained with a hand-rolled generalized MixIT loss on full mixes (no stems). Outputs feed a downstream pitch analyzer; the bar is usable partial separation, not clean stems.

This checkpoint (model.pt, from best.pt) was supervised-fine-tuned on synthetic ground truth (a procedural acid-techno generator with true stems), semi-supervised (interleaved real MixIT batches) to avoid forgetting real tracks. Measured against the blind baseline on held-out synthetic mixes: it breaks the blind identifiability bound where a melodic line is buried under a same-register pitched competitor (sep-RPA 0.89 → 0.96) while preserving real-track behaviour. See MELODIC_LINE.md in the source repo.

Use

from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("maulay/music-splitter-mixit", "model.pt")
# then: python separate.py --ckpt model.pt --in song.wav --out stems
#   or: python extract_melody.py --ckpt model.pt --in song.wav --out melody

Honest limitations

Partial separation (often concentrates energy into one output); same-register polyphony is the hard failure. The melodic-line pipeline measures a pitch line (CREPE) after separation — it does not separate instruments by pitch.