dubbing-demucs

The demucs checkpoint the Blaze dubbing pipeline separates speech with — 04573f0d, the vocals specialist from facebookresearch/demucs (MIT). Mirrored here so a Docker build pulls a pinned file instead of whatever torch.hub serves that day.

DEMUCS_MODEL=04573f0d in the service selects it. It is roughly 4x faster than htdemucs_ft, and the pipeline takes the background as mix-minus-vocals rather than the sum of the other stems, so the specialist's weak drums/bass/other cost nothing — measured within 0.1 dB of the original background, against 8.5 dB down when summing stems.

Branches: main is what builds pin. dev is for trying a different checkpoint first.