Hugging Face's logo

mozilla-ai
/
whisper-small-el

Automatic Speech Recognition
Transformers
TensorBoard
Safetensors
Greek
whisper
Eval Results
Model card Files Metrics Community
kostissz commited on
Commit
7ea5e86
·
verified ·
1 Parent(s): f1e07b7

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +20 -103
README.md CHANGED
@@ -1,118 +1,35 @@
1
  ---
2
- library_name: transformers
3
- license: apache-2.0
4
  base_model: openai/whisper-small
5
- tags:
6
- - generated_from_trainer
7
  datasets:
8
- - common_voice_17_0
9
- metrics:
10
- - wer
 
11
  model-index:
12
- - name: whisper-small-el
13
  results:
14
  - task:
15
- name: Automatic Speech Recognition
16
  type: automatic-speech-recognition
 
17
  dataset:
18
- name: common_voice_17_0
19
- type: common_voice_17_0
20
- config: el
21
- split: None
22
- args: el
23
  metrics:
24
- - name: Wer
25
- type: wer
26
- value: 30.64381658175081
27
  ---
28
 
29
- <!-- This model card has been generated automatically according to the information the Trainer had access to. You
30
- should probably proofread and complete it, then remove this comment. -->
31
-
32
- # whisper-small-el
33
-
34
- This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) on the common_voice_17_0 dataset.
35
- It achieves the following results on the evaluation set:
36
- - Loss: 0.3865
37
- - Model Preparation Time: 0.0041
38
- - Wer: 30.6438
39
-
40
- ## Model description
41
-
42
- More information needed
43
-
44
- ## Intended uses & limitations
45
-
46
- More information needed
47
-
48
- ## Training and evaluation data
49
-
50
- More information needed
51
-
52
- ## Training procedure
53
-
54
- ### Training hyperparameters
55
-
56
- The following hyperparameters were used during training:
57
- - learning_rate: 1e-05
58
- - train_batch_size: 24
59
- - eval_batch_size: 8
60
- - seed: 42
61
- - optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
62
- - lr_scheduler_type: linear
63
- - lr_scheduler_warmup_steps: 50
64
- - training_steps: 2000
65
- - mixed_precision_training: Native AMP
66
-
67
- ### Training results
68
 
69
- | Training Loss | Epoch | Step | Validation Loss | Model Preparation Time | Wer |
70
- |:-------------:|:-------:|:----:|:---------------:|:----------------------:|:-------:|
71
- | 0.5511 | 0.3311 | 50 | 0.3685 | 0.0041 | 39.3516 |
72
- | 0.3132 | 0.6623 | 100 | 0.3168 | 0.0041 | 35.6276 |
73
- | 0.2709 | 0.9934 | 150 | 0.2897 | 0.0041 | 33.4785 |
74
- | 0.1634 | 1.3245 | 200 | 0.2829 | 0.0041 | 33.1450 |
75
- | 0.1551 | 1.6556 | 250 | 0.2746 | 0.0041 | 32.5614 |
76
- | 0.1559 | 1.9868 | 300 | 0.2683 | 0.0041 | 31.8481 |
77
- | 0.0818 | 2.3179 | 350 | 0.2735 | 0.0041 | 31.3942 |
78
- | 0.0808 | 2.6490 | 400 | 0.2735 | 0.0041 | 31.9592 |
79
- | 0.0799 | 2.9801 | 450 | 0.2765 | 0.0041 | 32.4595 |
80
- | 0.0451 | 3.3113 | 500 | 0.2922 | 0.0041 | 31.4590 |
81
- | 0.0436 | 3.6424 | 550 | 0.2892 | 0.0041 | 31.0514 |
82
- | 0.0436 | 3.9735 | 600 | 0.2902 | 0.0041 | 31.3942 |
83
- | 0.0241 | 4.3046 | 650 | 0.3117 | 0.0041 | 31.2552 |
84
- | 0.0212 | 4.6358 | 700 | 0.3162 | 0.0041 | 31.0699 |
85
- | 0.0226 | 4.9669 | 750 | 0.3172 | 0.0041 | 30.8754 |
86
- | 0.0127 | 5.2980 | 800 | 0.3521 | 0.0041 | 32.5336 |
87
- | 0.0125 | 5.6291 | 850 | 0.3432 | 0.0041 | 31.1996 |
88
- | 0.0123 | 5.9603 | 900 | 0.3463 | 0.0041 | 31.4034 |
89
- | 0.0077 | 6.2914 | 950 | 0.3764 | 0.0041 | 31.0699 |
90
- | 0.0071 | 6.6225 | 1000 | 0.3607 | 0.0041 | 32.4317 |
91
- | 0.0062 | 6.9536 | 1050 | 0.3698 | 0.0041 | 30.8754 |
92
- | 0.0045 | 7.2848 | 1100 | 0.3758 | 0.0041 | 30.9588 |
93
- | 0.0035 | 7.6159 | 1150 | 0.3865 | 0.0041 | 30.6438 |
94
- | 0.0038 | 7.9470 | 1200 | 0.3856 | 0.0041 | 31.2830 |
95
- | 0.0027 | 8.2781 | 1250 | 0.3800 | 0.0041 | 30.8569 |
96
- | 0.0021 | 8.6093 | 1300 | 0.3858 | 0.0041 | 30.6901 |
97
- | 0.0022 | 8.9404 | 1350 | 0.3949 | 0.0041 | 31.1996 |
98
- | 0.0017 | 9.2715 | 1400 | 0.4020 | 0.0041 | 30.7920 |
99
- | 0.0016 | 9.6026 | 1450 | 0.4061 | 0.0041 | 30.9588 |
100
- | 0.0016 | 9.9338 | 1500 | 0.4111 | 0.0041 | 31.0514 |
101
- | 0.0014 | 10.2649 | 1550 | 0.4067 | 0.0041 | 31.1996 |
102
- | 0.0013 | 10.5960 | 1600 | 0.4093 | 0.0041 | 31.0144 |
103
- | 0.0013 | 10.9272 | 1650 | 0.4112 | 0.0041 | 30.8661 |
104
- | 0.0012 | 11.2583 | 1700 | 0.4126 | 0.0041 | 30.9680 |
105
- | 0.0012 | 11.5894 | 1750 | 0.4134 | 0.0041 | 30.9588 |
106
- | 0.0012 | 11.9205 | 1800 | 0.4145 | 0.0041 | 30.9217 |
107
- | 0.0011 | 12.2517 | 1850 | 0.4155 | 0.0041 | 30.8384 |
108
- | 0.0011 | 12.5828 | 1900 | 0.4160 | 0.0041 | 30.8939 |
109
- | 0.0011 | 12.9139 | 1950 | 0.4163 | 0.0041 | 30.8754 |
110
- | 0.0011 | 13.2450 | 2000 | 0.4164 | 0.0041 | 30.8754 |
111
 
 
112
 
113
- ### Framework versions
 
 
114
 
115
- - Transformers 4.48.0
116
- - Pytorch 2.5.1+cu124
117
- - Datasets 3.2.0
118
- - Tokenizers 0.21.0
 
1
  ---
 
 
2
  base_model: openai/whisper-small
 
 
3
  datasets:
4
+ - mozilla-foundation/common_voice_17_0
5
+ language: el
6
+ library_name: transformers
7
+ license: apache-2.0
8
  model-index:
9
+ - name: Finetuned openai/whisper-small on Greek
10
  results:
11
  - task:
 
12
  type: automatic-speech-recognition
13
+ name: Speech-to-Text
14
  dataset:
15
+ name: Common Voice (Greek)
16
+ type: common_voice
 
 
 
17
  metrics:
18
+ - type: wer
19
+ value: 30.644
 
20
  ---
21
 
22
+ # Finetuned openai/whisper-small on 3620 Greek training audio samples from mozilla-foundation/common_voice_17_0.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23
 
24
+ This model was created from the Mozilla.ai Blueprint:
25
+ [speech-to-text-finetune](https://github.com/mozilla-ai/speech-to-text-finetune).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
26
 
27
+ ## Evaluation results on 1701 audio samples of Greek:
28
 
29
+ ### Baseline model (before finetuning) on Greek
30
+ - Word Error Rate: 46.401
31
+ - Loss: 0.902
32
 
33
+ ### Finetuned model (after finetuning) on Greek
34
+ - Word Error Rate: 30.644
35
+ - Loss: 0.387