sammy786 commited on
Commit
b253da0
1 Parent(s): fadb14d

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +114 -0
README.md CHANGED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - hu
4
+ license: apache-2.0
5
+ tags:
6
+ - automatic-speech-recognition
7
+ - mozilla-foundation/common_voice_8_0
8
+ - generated_from_trainer
9
+ - fi
10
+ - robust-speech-event
11
+ - model_for_talk
12
+ datasets:
13
+ - mozilla-foundation/common_voice_8_0
14
+
15
+ model-index:
16
+ - name: sammy786/wav2vec2-xlsr-finnish
17
+ results:
18
+ - task:
19
+ name: Automatic Speech Recognition
20
+ type: automatic-speech-recognition
21
+ dataset:
22
+ name: Common Voice 8
23
+ type: mozilla-foundation/common_voice_8_0
24
+ args: fi
25
+ metrics:
26
+ - name: Test WER
27
+ type: wer
28
+ value: 0.2889
29
+ - name: Test CER
30
+ type: cer
31
+ value: 0.06
32
+ - task:
33
+ name: Automatic Speech Recognition
34
+ type: automatic-speech-recognition
35
+ dataset:
36
+ name: Robust Speech Event - Dev Data
37
+ type: speech-recognition-community-v2/dev_data
38
+ args: hu
39
+ metrics:
40
+ - name: Test WER
41
+ type: wer
42
+ value: 0.2889
43
+ - name: Test CER
44
+ type: cer
45
+ value: 0.06
46
+ ---
47
+ # sammy786/wav2vec2-xlsr-finnish
48
+
49
+ This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - fi dataset.
50
+ It achieves the following results on evaluation set (which is 10 percent of train data set merged with invalidated data, other, dev and validated datasets):
51
+ - Loss: 0.092132
52
+ - Wer: 0.226287
53
+
54
+ ## Model description
55
+ "facebook/wav2vec2-xls-r-300m" was finetuned.
56
+
57
+ ## Intended uses & limitations
58
+ More information needed
59
+ ## Training and evaluation data
60
+ Training data -
61
+ Common voice Finnish train.tsv, dev.tsv, invalidated.tsv and other.tsv
62
+
63
+ ## Training procedure
64
+ For creating the train dataset, all possible datasets were appended and 90-10 split was used.
65
+
66
+ ### Training hyperparameters
67
+
68
+ The following hyperparameters were used during training:
69
+
70
+ - learning_rate: 0.000095637994662983496
71
+ - train_batch_size: 32
72
+ - eval_batch_size: 16
73
+ - seed: 13
74
+ - gradient_accumulation_steps: 2
75
+ - total_train_batch_size: 64
76
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
77
+ - lr_scheduler_type: cosine_with_restarts
78
+ - lr_scheduler_warmup_steps: 500
79
+ - num_epochs: 80
80
+ - mixed_precision_training: Native AMP
81
+
82
+
83
+ ### Training results
84
+
85
+
86
+ | Step | Training Loss | Validation Loss | Wer |
87
+ |:-----:|:-------------:|:---------------:|:--------:|
88
+ | 1000 | 0.593200 | 0.180788 | 0.396133 |
89
+ | 2000 | 0.154900 | 0.112490 | 0.288999 |
90
+ | 3000 | 0.114900 | 0.103121 | 0.267442 |
91
+ | 4000 | 0.093900 | 0.102042 | 0.258950 |
92
+ | 5000 | 0.080700 | 0.100328 | 0.246407 |
93
+ | 6000 | 0.071100 | 0.097676 | 0.247322 |
94
+ | 7000 | 0.064300 | 0.091082 | 0.240267 |
95
+ | 8000 | 0.056300 | 0.094759 | 0.236086 |
96
+ | 9000 | 0.051300 | 0.092355 | 0.234387 |
97
+ | 10000 | 0.046800 | 0.093417 | 0.228377 |
98
+ | 11000 | 0.045000 | 0.092780 | 0.226810 |
99
+ | 12000 | 0.042100 | 0.092132 | 0.226287 |
100
+
101
+
102
+ ### Framework versions
103
+ - Transformers 4.16.0.dev0
104
+ - Pytorch 1.10.0+cu102
105
+ - Datasets 1.17.1.dev0
106
+ - Tokenizers 0.10.3
107
+
108
+ #### Evaluation Commands
109
+
110
+ 1. To evaluate on `mozilla-foundation/common_voice_8_0` with split `test`
111
+
112
+ ```bash
113
+ python eval.py --model_id sammy786/wav2vec2-xlsr-finnish --dataset mozilla-foundation/common_voice_8_0 --config fi --split test
114
+ ```