alexwengg commited on
Commit
7eab74f
·
verified ·
1 Parent(s): df2625a

Document provenance and licenses of the legacy segmentation/wespeaker artifacts

Browse files
Files changed (2) hide show
  1. NOTICE.md +46 -4
  2. README.md +5 -3
NOTICE.md CHANGED
@@ -29,13 +29,55 @@ Appropriate attribution should identify pyannote, WeSpeaker, BUT Speech@FIT,
29
  and Fluid Inference, retain the citations in README.md, link CC-BY-4.0, and
30
  indicate that the files are modified Core ML conversions.
31
 
32
- ## Legacy exclusions
33
 
34
  `pyannote_segmentation.mlmodelc`, `wespeaker.mlmodelc`,
35
  `wespeaker_v2.mlmodelc`, `wespeaker_int8.mlmodelc`, and their packages predate
36
- the supported Community-1 artifact set. They are retained for compatibility but
37
- are not covered by this Community-1 provenance and license-scope confirmation.
38
- Their original source and licensing must be evaluated separately.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
39
 
40
  Fluid Inference can grant rights only to the extent it is authorized to do so.
41
  This notice does not restrict or replace rights granted directly by upstream
 
29
  and Fluid Inference, retain the citations in README.md, link CC-BY-4.0, and
30
  indicate that the files are modified Core ML conversions.
31
 
32
+ ## Legacy artifacts
33
 
34
  `pyannote_segmentation.mlmodelc`, `wespeaker.mlmodelc`,
35
  `wespeaker_v2.mlmodelc`, `wespeaker_int8.mlmodelc`, and their packages predate
36
+ the Community-1 artifact set and are not covered by the Community-1 scope
37
+ above. They are Core ML conversions of the pyannote/speaker-diarization-3.1
38
+ components, under their own upstream licenses:
39
+
40
+ | Artifact | Upstream checkpoint | License |
41
+ |---|---|---|
42
+ | `pyannote_segmentation` | [`pyannote/segmentation-3.0`](https://huggingface.co/pyannote/segmentation-3.0) @ `e66f3d3b` (Copyright (c) 2023 CNRS) | MIT |
43
+ | `wespeaker` (fp32) | [`pyannote/wespeaker-voxceleb-resnet34-LM`](https://huggingface.co/pyannote/wespeaker-voxceleb-resnet34-LM) @ `837717dd`, wrapping WeSpeaker's VoxCeleb ResNet34-LM | CC-BY-4.0 |
44
+ | `wespeaker_v2`, `wespeaker_int8` (byte-identical; 8-bit palettized weights) | same as `wespeaker` | CC-BY-4.0 |
45
+
46
+ The original conversion scripts were not preserved, so this was confirmed by
47
+ output parity on AMI ES2004a against the PyTorch checkpoints above (pyannote.audio
48
+ 3.3.2, CPU): segmentation frame-level argmax agreement 99.98%; speaker
49
+ embeddings cosine 1.0000 (`wespeaker`) and >= 0.997 (`wespeaker_v2`/`wespeaker_int8`).
50
+
51
+ Fluid Inference converted these checkpoints to Core ML with fixed input shapes,
52
+ fused the fbank frontend into the embedding model, and palettized the
53
+ `wespeaker_v2`/`wespeaker_int8` weights to 8 bits. Attribution for the
54
+ embedding models should identify WeSpeaker, pyannote, and Fluid Inference,
55
+ link CC-BY-4.0, and indicate that the files are modified Core ML conversions.
56
+ The segmentation model's MIT notice is reproduced below.
57
+
58
+ ```
59
+ MIT License
60
+
61
+ Copyright (c) 2023 CNRS
62
+
63
+ Permission is hereby granted, free of charge, to any person obtaining a copy
64
+ of this software and associated documentation files (the "Software"), to deal
65
+ in the Software without restriction, including without limitation the rights
66
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
67
+ copies of the Software, and to permit persons to whom the Software is
68
+ furnished to do so, subject to the following conditions:
69
+
70
+ The above copyright notice and this permission notice shall be included in all
71
+ copies or substantial portions of the Software.
72
+
73
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
74
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
75
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
76
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
77
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
78
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
79
+ SOFTWARE.
80
+ ```
81
 
82
  Fluid Inference can grant rights only to the extent it is authorized to do so.
83
  This notice does not restrict or replace rights granted directly by upstream
README.md CHANGED
@@ -63,9 +63,11 @@ build environment. It is available at
63
 
64
  The repository also retains `pyannote_segmentation.mlmodelc`,
65
  `wespeaker.mlmodelc`, `wespeaker_v2.mlmodelc`, and `wespeaker_int8.mlmodelc`
66
- for FluidAudio's legacy online diarizer. They predate the Community-1 export,
67
- record a different toolchain, and are not included in the Community-1
68
- provenance or license-scope confirmation in `NOTICE.md`.
 
 
69
 
70
  ## Technical specifications
71
 
 
63
 
64
  The repository also retains `pyannote_segmentation.mlmodelc`,
65
  `wespeaker.mlmodelc`, `wespeaker_v2.mlmodelc`, and `wespeaker_int8.mlmodelc`
66
+ for FluidAudio's legacy online diarizer. They predate the Community-1 export
67
+ and are not covered by its license scope. They are Core ML conversions of
68
+ [`pyannote/segmentation-3.0`](https://huggingface.co/pyannote/segmentation-3.0)
69
+ (MIT) and [`pyannote/wespeaker-voxceleb-resnet34-LM`](https://huggingface.co/pyannote/wespeaker-voxceleb-resnet34-LM)
70
+ (CC-BY-4.0); see `NOTICE.md` for revisions, parity evidence, and attribution.
71
 
72
  ## Technical specifications
73