Hugging Face's logo

razhan
/
whisper-small-me

Automatic Speech Recognition
Transformers
Safetensors
whisper
Generated from Trainer
Model card Files
xet
Community
razhan commited on
Commit
03afe9f
·
verified ·
1 Parent(s): 1a18fe5

Model save

Browse files
Files changed (2) hide show
  1. README.md +30 -37
  2. generation_config.json +0 -1
README.md CHANGED
@@ -16,35 +16,23 @@ should probably proofread and complete it, then remove this comment. -->
16
 
17
  This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) on the None dataset.
18
  It achieves the following results on the evaluation set:
19
- - Loss: 1.4680
20
- - Mazanderani Bleu: 0.0
21
- - Mazanderani Chrf: 0.0
22
- - Mazanderani Wer: 0.9419
23
- - Mazanderani Cer: 0.4586
24
- - Gilaki Bleu: 0.0
25
- - Gilaki Chrf: 0.0
26
- - Gilaki Wer: 0.9789
27
- - Gilaki Cer: 0.4864
28
- - Zazaki Bleu: 0.0
29
- - Zazaki Chrf: 0.0
30
- - Zazaki Wer: 1.0074
31
- - Zazaki Cer: 0.8905
32
- - Laki Kurdish Bleu: 0.0
33
- - Laki Kurdish Chrf: 0.0
34
- - Laki Kurdish Wer: 0.6814
35
- - Laki Kurdish Cer: 0.2116
36
- - Talysh Bleu: 0.0
37
- - Talysh Chrf: 0.0
38
  - Talysh Wer: 1.0
39
  - Talysh Cer: 0.5
40
- - Hawrami Bleu: 0.0
41
- - Hawrami Chrf: 0.0
42
- - Hawrami Wer: 0.3160
43
- - Hawrami Cer: 0.0656
44
- - Southern Kurdish Bleu: 0.0
45
- - Southern Kurdish Chrf: 0.0
46
- - Southern Kurdish Wer: 0.6534
47
- - Southern Kurdish Cer: 0.2319
48
 
49
  ## Model description
50
 
@@ -64,23 +52,28 @@ More information needed
64
 
65
  The following hyperparameters were used during training:
66
  - learning_rate: 1e-05
67
- - train_batch_size: 64
68
- - eval_batch_size: 32
69
  - seed: 42
 
 
 
 
70
  - optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
71
  - lr_scheduler_type: linear
72
- - lr_scheduler_warmup_steps: 10
73
- - num_epochs: 4.0
74
  - mixed_precision_training: Native AMP
75
 
76
  ### Training results
77
 
78
- | Training Loss | Epoch | Step | Validation Loss | Mazanderani Bleu | Mazanderani Chrf | Mazanderani Wer | Mazanderani Cer | Gilaki Bleu | Gilaki Chrf | Gilaki Wer | Gilaki Cer | Zazaki Bleu | Zazaki Chrf | Zazaki Wer | Zazaki Cer | Laki Kurdish Bleu | Laki Kurdish Chrf | Laki Kurdish Wer | Laki Kurdish Cer | Talysh Bleu | Talysh Chrf | Talysh Wer | Talysh Cer | Hawrami Bleu | Hawrami Chrf | Hawrami Wer | Hawrami Cer | Southern Kurdish Bleu | Southern Kurdish Chrf | Southern Kurdish Wer | Southern Kurdish Cer |
79
- |:-------------:|:-----:|:----:|:---------------:|:----------------:|:----------------:|:---------------:|:---------------:|:-----------:|:-----------:|:----------:|:----------:|:-----------:|:-----------:|:----------:|:----------:|:-----------------:|:-----------------:|:----------------:|:----------------:|:-----------:|:-----------:|:----------:|:----------:|:------------:|:------------:|:-----------:|:-----------:|:---------------------:|:---------------------:|:--------------------:|:--------------------:|
80
- | 1.1637 | 1.0 | 653 | 1.5382 | 0.0 | 0.0 | 0.9516 | 0.4696 | 0.0 | 0.0 | 0.9719 | 0.4789 | 0.0 | 0.0 | 1.0147 | 0.8900 | 0.0 | 0.0 | 0.7374 | 0.2241 | 0.0 | 0.0 | 1.0 | 0.5 | 0.0 | 0.0 | 0.3587 | 0.0754 | 0.0 | 0.0 | 0.6440 | 0.2060 |
81
- | 0.9336 | 2.0 | 1306 | 1.4697 | 0.0 | 0.0 | 0.9485 | 0.4687 | 0.0 | 0.0 | 0.9729 | 0.4833 | 0.0 | 0.0 | 1.0098 | 0.8910 | 0.0 | 0.0 | 0.7051 | 0.2182 | 0.0 | 0.0 | 1.0 | 0.5 | 0.0 | 0.0 | 0.3318 | 0.0690 | 0.0 | 0.0 | 0.6282 | 0.2022 |
82
- | 0.4965 | 3.0 | 1959 | 1.4548 | 0.0 | 0.0 | 0.9440 | 0.4593 | 0.0 | 0.0 | 0.9794 | 0.4867 | 0.0 | 0.0 | 1.0049 | 0.8941 | 0.0 | 0.0 | 0.6904 | 0.2173 | 0.0 | 0.0 | 1.0 | 0.5 | 0.0 | 0.0 | 0.3227 | 0.0681 | 0.0 | 0.0 | 0.6381 | 0.2090 |
83
- | 0.5566 | 4.0 | 2612 | 1.4680 | 0.0 | 0.0 | 0.9419 | 0.4586 | 0.0 | 0.0 | 0.9789 | 0.4864 | 0.0 | 0.0 | 1.0074 | 0.8905 | 0.0 | 0.0 | 0.6814 | 0.2116 | 0.0 | 0.0 | 1.0 | 0.5 | 0.0 | 0.0 | 0.3160 | 0.0656 | 0.0 | 0.0 | 0.6534 | 0.2319 |
 
84
 
85
 
86
  ### Framework versions
 
16
 
17
  This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) on the None dataset.
18
  It achieves the following results on the evaluation set:
19
+ - Loss: 0.4509
20
+ - Mazanderani Wer: 0.5278
21
+ - Mazanderani Cer: 0.1615
22
+ - Gilaki Wer: 0.9176
23
+ - Gilaki Cer: 0.3175
24
+ - Zazaki Wer: 0.6324
25
+ - Zazaki Cer: 0.2020
26
+ - Laki Kurdish Wer: 0.4904
27
+ - Laki Kurdish Cer: 0.1247
 
 
 
 
 
 
 
 
 
 
28
  - Talysh Wer: 1.0
29
  - Talysh Cer: 0.5
30
+ - Hawrami Wer: 0.3459
31
+ - Hawrami Cer: 0.0708
32
+ - Southern Kurdish Wer: 0.4930
33
+ - Southern Kurdish Cer: 0.1718
34
+ - Avg Wer: 0.6296
35
+ - Avg Cer: 0.2212
 
 
36
 
37
  ## Model description
38
 
 
52
 
53
  The following hyperparameters were used during training:
54
  - learning_rate: 1e-05
55
+ - train_batch_size: 128
56
+ - eval_batch_size: 128
57
  - seed: 42
58
+ - distributed_type: multi-GPU
59
+ - num_devices: 2
60
+ - total_train_batch_size: 256
61
+ - total_eval_batch_size: 256
62
  - optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
63
  - lr_scheduler_type: linear
64
+ - lr_scheduler_warmup_steps: 100
65
+ - num_epochs: 5.0
66
  - mixed_precision_training: Native AMP
67
 
68
  ### Training results
69
 
70
+ | Training Loss | Epoch | Step | Validation Loss | Mazanderani Wer | Mazanderani Cer | Gilaki Wer | Gilaki Cer | Zazaki Wer | Zazaki Cer | Laki Kurdish Wer | Laki Kurdish Cer | Talysh Wer | Talysh Cer | Hawrami Wer | Hawrami Cer | Southern Kurdish Wer | Southern Kurdish Cer | Avg Wer | Avg Cer |
71
+ |:-------------:|:-----:|:----:|:---------------:|:---------------:|:---------------:|:----------:|:----------:|:----------:|:----------:|:----------------:|:----------------:|:----------:|:----------:|:-----------:|:-----------:|:--------------------:|:--------------------:|:-------:|:-------:|
72
+ | 0.6946 | 1.0 | 82 | 0.7662 | 0.8120 | 0.3224 | 0.9556 | 0.3751 | 1.0270 | 1.0389 | 0.7939 | 0.2733 | 0.9167 | 0.3333 | 0.5246 | 0.1155 | 0.6801 | 0.2192 | 0.8157 | 0.3825 |
73
+ | 0.3864 | 2.0 | 164 | 0.5172 | 0.6093 | 0.1846 | 0.9579 | 0.3454 | 0.7451 | 0.2813 | 0.5644 | 0.1547 | 1.0 | 0.5 | 0.3806 | 0.0773 | 0.5484 | 0.1743 | 0.6865 | 0.2454 |
74
+ | 0.3234 | 3.0 | 246 | 0.4686 | 0.5476 | 0.1652 | 0.9633 | 0.3514 | 0.6740 | 0.2363 | 0.5231 | 0.1362 | 1.0 | 0.5 | 0.3533 | 0.0722 | 0.4936 | 0.1674 | 0.6507 | 0.2327 |
75
+ | 0.2786 | 4.0 | 328 | 0.4527 | 0.5278 | 0.1596 | 0.9083 | 0.3179 | 0.6422 | 0.2133 | 0.5047 | 0.1334 | 1.0 | 0.5 | 0.3485 | 0.0732 | 0.4948 | 0.1769 | 0.6323 | 0.2249 |
76
+ | 0.2586 | 5.0 | 410 | 0.4509 | 0.5278 | 0.1615 | 0.9176 | 0.3175 | 0.6324 | 0.2020 | 0.4904 | 0.1247 | 1.0 | 0.5 | 0.3459 | 0.0708 | 0.4930 | 0.1718 | 0.6296 | 0.2212 |
77
 
78
 
79
  ### Framework versions
generation_config.json CHANGED
@@ -150,7 +150,6 @@
150
  "<|yo|>": 50325,
151
  "<|zh|>": 50260
152
  },
153
- "language": "persian",
154
  "max_initial_timestamp_index": 50,
155
  "max_length": 448,
156
  "no_timestamps_token_id": 50363,
 
150
  "<|yo|>": 50325,
151
  "<|zh|>": 50260
152
  },
 
153
  "max_initial_timestamp_index": 50,
154
  "max_length": 448,
155
  "no_timestamps_token_id": 50363,