source-separation
senselab
mirror
File size: 2,296 Bytes
92c223d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
---
license: other
license_name: unresolved-upstream
tags: [source-separation, senselab, mirror]
---

# 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](https://arxiv.org/abs/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

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