Hugging Face's logo

BUT-FIT
/
diarizen-wavlm-large-s80-md

Voice Activity Detection
Transformers
PyTorch
speaker
speaker-diarization
meeting
wavlm
wespeaker
diarizen
pyannote
pyannote-audio-pipeline
Model card Files
xet
 Community
6
Jyhan003 commited on
Commit
a67e6ba
·
1 Parent(s): 75f8b09

restore original model weights for backward compatibility

Browse files
Files changed (3) hide show
  1. README.md +9 -9
  2. config.toml +0 -2
  3. pytorch_model.bin +2 -2
README.md CHANGED
@@ -18,8 +18,9 @@ This hub features the pre-trained model by [DiariZen](https://github.com/BUTSpee
18
 
19
  Then structured pruning at 80% sparsity is applied. After pruning, the number of parameters in WavLM Large is reduced from **316.6M to 63.3M**, and the computational cost (MACs) decreases from **17.8G to 3.8G** per second. When loading this model, please ensure **non-commercial** usage, in accordance with the CC BY-NC 4.0 license.
20
 
21
- ## Note
22
- This repository contains the latest version of DiariZen, which supports up to four overlapping speakers for improved diarization performance. Please make sure to pull the most recent code before using the model, as outdated versions may lead to unexpected issues or degraded results. If you prefer to use the original (legacy) model, you can find it [here](https://huggingface.co/BUT-FIT/diarizen-wavlm-large-s80-md-origin).
 
23
 
24
  ## Usage
25
  ```python
@@ -46,15 +47,14 @@ diar_results = diar_pipeline('audio.wav', sess_name='session_name')
46
  ## Results (collar=0s)
47
  | Dataset | [Pyannote v3.1](https://github.com/pyannote/pyannote-audio) | DiariZen |
48
  |:---------------|:-----------:|:-----------:|
49
- | AMI | 22.4 | 13.9 |
50
- | AISHELL-4 | 12.2 | 10.1 |
51
- | AliMeeting | 24.4 | 10.8 |
52
- | NOTSOFAR-1 | - | 16.7 |
53
- | MSDWild | 25.3 | 15.8 |
54
  | DIHARD3 | 21.7 | 14.5 |
55
  | RAMC | 22.2 | 11.0 |
56
- | VoxConverse | 11.3 | 9.1 |
57
-
58
  ## Citation
59
  If you found this work helpful, please consider citing:
60
  ```
 
18
 
19
  Then structured pruning at 80% sparsity is applied. After pruning, the number of parameters in WavLM Large is reduced from **316.6M to 63.3M**, and the computational cost (MACs) decreases from **17.8G to 3.8G** per second. When loading this model, please ensure **non-commercial** usage, in accordance with the CC BY-NC 4.0 license.
20
 
21
+ ## Note:
22
+ This is the original version of the repository. It is optimized for cases where no more than 3 speakers overlap at any given time. For scenarios with high speaker overlap (up to 4 simultaneous speakers), the [new repository](https://huggingface.co/BUT-FIT/diarizen-wavlm-large-s80-md-v2) is recommended as it provides significantly better performance under such conditions.
23
+
24
 
25
  ## Usage
26
  ```python
 
47
  ## Results (collar=0s)
48
  | Dataset | [Pyannote v3.1](https://github.com/pyannote/pyannote-audio) | DiariZen |
49
  |:---------------|:-----------:|:-----------:|
50
+ | AMI | 22.4 | 14.0 |
51
+ | AISHELL-4 | 12.2 | 9.8 |
52
+ | AliMeeting | 24.4 | 12.5 |
53
+ | NOTSOFAR-1 | - | 17.9 |
54
+ | MSDWild | 25.3 | 15.6 |
55
  | DIHARD3 | 21.7 | 14.5 |
56
  | RAMC | 22.2 | 11.0 |
57
+ | VoxConverse | 11.3 | 9.2 |
 
58
  ## Citation
59
  If you found this work helpful, please consider citing:
60
  ```
config.toml CHANGED
@@ -10,8 +10,6 @@ ffn_hidden = 1024
10
  num_head = 4
11
  num_layer = 4
12
  dropout = 0.1
13
- max_speakers_per_chunk = 4
14
- max_speakers_per_frame = 4
15
  chunk_size = 16
16
  use_posi = false
17
  output_activate_function = false
 
10
  num_head = 4
11
  num_layer = 4
12
  dropout = 0.1
 
 
13
  chunk_size = 16
14
  use_posi = false
15
  output_activate_function = false
pytorch_model.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:689a10f0cbde6345b4bf9b1670cf3196488634b34e5b07bea0fe40d508e1196b
3
- size 278071526
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0252b5d926600ad434778a85314f3164601914d36d355cbad3e7e30be968a675
3
+ size 278066342