GIanlucaRub commited on
Commit
83b51a1
1 Parent(s): f657b5b

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +85 -0
README.md ADDED
@@ -0,0 +1,85 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - it
4
+ license: apache-2.0
5
+ tags:
6
+ - hf-asr-leaderboard
7
+ - generated_from_trainer
8
+ datasets:
9
+ - mozilla-foundation/common_voice_11_0
10
+ metrics:
11
+ - wer
12
+ model-index:
13
+ - name: Whisper Tiny It 3 - Gianluca Ruberto
14
+ results:
15
+ - task:
16
+ name: Automatic Speech Recognition
17
+ type: automatic-speech-recognition
18
+ dataset:
19
+ name: Common Voice 11.0
20
+ type: mozilla-foundation/common_voice_11_0
21
+ config: it
22
+ split: test[:10%]
23
+ args: 'config: hi, split: test'
24
+ metrics:
25
+ - name: Wer
26
+ type: wer
27
+ value: 43.233499722684414
28
+ ---
29
+
30
+
31
+ # Whisper Tiny It 3 - Gianluca Ruberto
32
+
33
+ This model is a fine-tuned version of [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) on the Common Voice 11.0 dataset.
34
+ It achieves the following results on the evaluation set:
35
+ - Loss: 0.711673
36
+ - Wer: 43.233500
37
+
38
+ ## Model description
39
+
40
+ This model is the openai whisper small transformer adapted for Italian audio to text transcription. This model has weight decay set to 0.1 to cope with overfitting.
41
+
42
+ ## Intended uses & limitations
43
+
44
+ The model is available through its [HuggingFace web app](https://huggingface.co/spaces/GIanlucaRub/whisper-it)
45
+
46
+ ## Training and evaluation data
47
+
48
+ Data used for training is the initial 10% of train and validation of [Italian Common Voice](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0/viewer/it/train) 11.0 from Mozilla Foundation.
49
+ The dataset used for evaluation is the initial 10% of test of Italian Common Voice.
50
+ Weight decay showed to have slightly better result also on the evaluation dataset.
51
+
52
+ ## Training procedure
53
+
54
+ After loading the pre trained model, it has been trained on the dataset.
55
+
56
+ ### Training hyperparameters
57
+
58
+ The following hyperparameters were used during training:
59
+ - learning_rate: 1e-05
60
+ - train_batch_size: 16
61
+ - eval_batch_size: 8
62
+ - seed: 42
63
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
64
+ - lr_scheduler_type: linear
65
+ - lr_scheduler_warmup_steps: 500
66
+ - training_steps: 4000
67
+ - mixed_precision_training: Native AMP
68
+ - weight_decay: 0.1
69
+
70
+ ### Training results
71
+
72
+ | Training Loss | Epoch | Step | Validation Loss | Wer |
73
+ |:-------------:|:-----:|:----:|:---------------:|:-------:|
74
+ | 0.5837 | 0.95 | 1000 | 0.790374 | 50.2981 |
75
+ | 0.4183 | 1.91 | 2000 | 0.730100 | 45.4174 |
76
+ | 0.3147 | 2.86 | 3000 | 0.713152 | 44.3150 |
77
+ | 0.2670 | 3.82 | 4000 | 0.711673 | 43.2335 |
78
+
79
+
80
+ ### Framework versions
81
+
82
+ - Transformers 4.26.0.dev0
83
+ - Pytorch 1.12.1+cu113
84
+ - Datasets 2.7.1
85
+ - Tokenizers 0.13.2