Document provenance and licenses of the legacy segmentation/wespeaker artifacts
Browse files
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
|
| 33 |
|
| 34 |
`pyannote_segmentation.mlmodelc`, `wespeaker.mlmodelc`,
|
| 35 |
`wespeaker_v2.mlmodelc`, `wespeaker_int8.mlmodelc`, and their packages predate
|
| 36 |
-
the
|
| 37 |
-
|
| 38 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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 |
-
|
| 68 |
-
|
|
|
|
|
|
|
| 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 |
|