Hugging Face's logo

sammy786
/
wav2vec2-xlsr-finnish

Automatic Speech Recognition
Transformers
PyTorch
Finnish
wav2vec2
Generated from Trainer
hf-asr-leaderboard
model_for_talk
mozilla-foundation/common_voice_8_0
robust-speech-event
Eval Results
Inference Endpoints
Model card Files Community
sammy786 commited on
Commit
f7d585c
•
1 Parent(s): f73e79e

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +51 -27
README.md CHANGED
@@ -25,10 +25,10 @@ model-index:
25
  metrics:
26
  - name: Test WER
27
  type: wer
28
- value: 28.89
29
  - name: Test CER
30
  type: cer
31
- value: 6.23
32
  - task:
33
  name: Automatic Speech Recognition
34
  type: automatic-speech-recognition
@@ -39,20 +39,20 @@ model-index:
39
  metrics:
40
  - name: Test WER
41
  type: wer
42
- value: 28.89
43
  - name: Test CER
44
  type: cer
45
- value: 6.23
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: 9.2132
52
- - Wer: 22.6287
53
 
54
  ## Model description
55
- "facebook/wav2vec2-xls-r-300m" was finetuned.
56
 
57
  ## Intended uses & limitations
58
  More information needed
@@ -67,36 +67,60 @@ For creating the train dataset, all possible datasets were appended and 90-10 sp
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
 
25
  metrics:
26
  - name: Test WER
27
  type: wer
28
+ value: 13.341
29
  - name: Test CER
30
  type: cer
31
+ value: 3.23
32
  - task:
33
  name: Automatic Speech Recognition
34
  type: automatic-speech-recognition
 
39
  metrics:
40
  - name: Test WER
41
  type: wer
42
+ value: 13.341
43
  - name: Test CER
44
  type: cer
45
+ value: 3.23
46
  ---
47
  # sammy786/wav2vec2-xlsr-finnish
48
 
49
+ This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) 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: 8.7555
52
+ - Wer: 23.0231
53
 
54
  ## Model description
55
+ "facebook/wav2vec2-xls-r-1b" was finetuned.
56
 
57
  ## Intended uses & limitations
58
  More information needed
 
67
 
68
  The following hyperparameters were used during training:
69
 
70
+ - learning_rate: 0.000045637994662983496
71
+ - train_batch_size: 8
72
  - eval_batch_size: 16
73
  - seed: 13
74
+ - gradient_accumulation_steps: 4
75
+ - total_train_batch_size: 32
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: 30
80
  - mixed_precision_training: Native AMP
81
 
82
 
83
  ### Training results
84
 
85
 
86
+ | Step | Training Loss | Validation Loss | Wer |
87
+ |:----:|:-------------:|:---------------:|:--------:|
88
+ | 200 | 4.253700 | 0.881733 | 0.967007 |
89
+ | 400 | 0.864800 | 0.226977 | 0.420836 |
90
+ | 600 | 0.607000 | 0.157473 | 0.343375 |
91
+ | 800 | 0.380200 | 0.145640 | 0.302672 |
92
+ | 1000 | 0.318400 | 0.128028 | 0.293886 |
93
+ | 1200 | 0.261100 | 0.121414 | 0.289941 |
94
+ | 1400 | 0.232300 | 0.113451 | 0.279182 |
95
+ | 1600 | 0.216600 | 0.113649 | 0.282948 |
96
+ | 1800 | 0.202500 | 0.112375 | 0.276134 |
97
+ | 2000 | 0.190000 | 0.105725 | 0.273803 |
98
+ | 2200 | 0.171000 | 0.109715 | 0.270755 |
99
+ | 2400 | 0.156500 | 0.105042 | 0.264300 |
100
+ | 2600 | 0.155600 | 0.108337 | 0.260714 |
101
+ | 2800 | 0.149100 | 0.112435 | 0.263583 |
102
+ | 3000 | 0.145100 | 0.106193 | 0.261969 |
103
+ | 3200 | 0.131700 | 0.102860 | 0.251210 |
104
+ | 3400 | 0.129100 | 0.096058 | 0.246907 |
105
+ | 3600 | 0.121600 | 0.099932 | 0.246369 |
106
+ | 3800 | 0.112000 | 0.099041 | 0.244397 |
107
+ | 4000 | 0.114100 | 0.101566 | 0.242604 |
108
+ | 4200 | 0.111500 | 0.089498 | 0.239197 |
109
+ | 4400 | 0.099800 | 0.092835 | 0.240990 |
110
+ | 4600 | 0.095300 | 0.093518 | 0.238121 |
111
+ | 4800 | 0.094300 | 0.090783 | 0.240631 |
112
+ | 5000 | 0.089000 | 0.094046 | 0.238479 |
113
+ | 5200 | 0.088000 | 0.089342 | 0.235252 |
114
+ | 5400 | 0.083600 | 0.087770 | 0.234535 |
115
+ | 5600 | 0.083600 | 0.088804 | 0.234355 |
116
+ | 5800 | 0.080300 | 0.090168 | 0.231307 |
117
+ | 6000 | 0.078100 | 0.090163 | 0.230949 |
118
+ | 6200 | 0.075600 | 0.088876 | 0.232383 |
119
+ | 6400 | 0.078700 | 0.087235 | 0.232024 |
120
+ | 6600 | 0.074800 | 0.086825 | 0.231486 |
121
+ | 6800 | 0.076400 | 0.087308 | 0.231845 |
122
+ | 7000 | 0.070700 | 0.087695 | 0.230769 |
123
+ | 7200 | 0.075500 | 0.087555 | 0.230231 |
124
 
125
 
126
  ### Framework versions