sensein/unasdiff-diffusion-priors

unasdiff diffusion source priors (mirror)

Mirror of the two released priors for Unsupervised Single-Channel Audio Separation with Diffusion Source Priors (Shi et al., AAAI 2026), arXiv:2512.07226.

Held so that senselab's source_separation unasdiff backend can pin a revision and a content hash. Nothing here is original work of the Senseable Intelligence Group.

Contents

File	What it is
speech_source.pt	Speech prior, trained on VCTK. num_class = 1; its only label is 0.
sound_source.pt	General-sound prior, trained on FSDKaggle2018. num_class = 50, of which indices 0–40 are the 41 trained classes; the remaining nine slots are unused headroom.
atten_unet_vctk.toml	Upstream config for the speech prior.
atten_unet_fsd.toml	Upstream config for the sound prior.

The two label spaces are separate. A label index is only meaningful against the prior it belongs to.

Each checkpoint carries model and ema state dicts. Upstream's own loader returns the ema copy; loading model instead runs but separates measurably worse, without failing.

Provenance

• Upstream code: https://github.com/RunwuShi/unasdiff, pinned at 5a5d70cdc94fe9d034892a1c5bc68ad1a67d2daa.
• Weights: the authors' own Google Drive releases linked from the upstream README.

Licence status — unresolved

The upstream repository has no LICENSE file and no licence statement, so the default is all rights reserved. A request for an explicit licence has been opened upstream; this mirror stays private until it is answered.

Known limitation, from the authors

The source-model-based separation approach is not well suited for same-class source separation (e.g. speech separation), because it lacks speaker-conditioning.

senselab exposes speech–speech separation anyway, with that caveat in the docstring, because the alternative is a user rediscovering it by measurement.

Citation

@inproceedings{shi2026unasdiff,
  title     = {Unsupervised Single-Channel Audio Separation with Diffusion Source Priors},
  booktitle = {Proc. AAAI},
  year      = {2026}
}