Hugging Face's logo

comodoro
/
wav2vec2-xls-r-300m-cs-250

Automatic Speech Recognition
Transformers
PyTorch
Safetensors
Czech
wav2vec2
Generated from Trainer
hf-asr-leaderboard
mozilla-foundation/common_voice_8_0
robust-speech-event
xlsr-fine-tuning-week
Eval Results
Inference Endpoints
Model card Files Community
2
comodoro commited on
Commit
e1bad10
•
1 Parent(s): 3ad836a

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +48 -16
README.md CHANGED
@@ -26,10 +26,10 @@ model-index:
26
  metrics:
27
  - name: Test WER
28
  type: wer
29
- value: 10.0
30
  - name: Test CER
31
  type: cer
32
- value: 2.6
33
  ---
34
 
35
  # Czech wav2vec2-xls-r-300m-cs-250
@@ -37,18 +37,13 @@ model-index:
37
  This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the common_voice 8.0 dataset as well as other datasets listed below.
38
 
39
  It achieves the following results on the evaluation set:
40
- - eval_loss: 0.1304
41
- - eval_wer: 0.1517
42
- - eval_cer: 0.0326
43
- - eval_runtime: 358.9895
44
- - eval_samples_per_second: 20.243
45
- - eval_steps_per_second: 2.532
46
- - epoch: 3.13
47
- - step: 31200
48
 
49
  The `eval.py` script results using a LM are:
50
- WER: 0.10053685691079459
51
- CER: 0.025859623842234124
52
 
53
  ## Model description
54
 
@@ -107,19 +102,56 @@ The Common Voice 8.0 `train` and `validation` datasets were used for training, a
107
 
108
  - Plátek, Ondřej; Dušek, Ondřej and Jurčíček, Filip, 2016, Vystadial 2016 – Czech data, LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles University, http://hdl.handle.net/11234/1-1740.
109
 
 
110
  ### Training hyperparameters
111
 
112
  The following hyperparameters were used during training:
113
- - learning_rate: 1e-05
114
- - train_batch_size: 16
115
  - eval_batch_size: 8
116
  - seed: 42
117
  - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
118
  - lr_scheduler_type: linear
119
- - lr_scheduler_warmup_steps: 600
120
- - num_epochs: 50
121
  - mixed_precision_training: Native AMP
122
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
123
  ### Framework versions
124
 
125
  - Transformers 4.16.2
 
26
  metrics:
27
  - name: Test WER
28
  type: wer
29
+ value: 7.3
30
  - name: Test CER
31
  type: cer
32
+ value: 2.1
33
  ---
34
 
35
  # Czech wav2vec2-xls-r-300m-cs-250
 
37
  This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the common_voice 8.0 dataset as well as other datasets listed below.
38
 
39
  It achieves the following results on the evaluation set:
40
+ - Loss: 0.1271
41
+ - Wer: 0.1475
42
+ - Cer: 0.0329
 
 
 
 
 
43
 
44
  The `eval.py` script results using a LM are:
45
+ - WER: 0.07274312090176113
46
+ - CER: 0.021207369275558875
47
 
48
  ## Model description
49
 
 
102
 
103
  - Plátek, Ondřej; Dušek, Ondřej and Jurčíček, Filip, 2016, Vystadial 2016 – Czech data, LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles University, http://hdl.handle.net/11234/1-1740.
104
 
105
+
106
  ### Training hyperparameters
107
 
108
  The following hyperparameters were used during training:
109
+ - learning_rate: 0.0001
110
+ - train_batch_size: 32
111
  - eval_batch_size: 8
112
  - seed: 42
113
  - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
114
  - lr_scheduler_type: linear
115
+ - lr_scheduler_warmup_steps: 800
116
+ - num_epochs: 5
117
  - mixed_precision_training: Native AMP
118
 
119
+ ### Training results
120
+
121
+ | Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
122
+ |:-------------:|:-----:|:-----:|:---------------:|:------:|:------:|
123
+ | 3.4203 | 0.16 | 800 | 3.3148 | 1.0 | 1.0 |
124
+ | 2.8151 | 0.32 | 1600 | 0.8508 | 0.8938 | 0.2345 |
125
+ | 0.9411 | 0.48 | 2400 | 0.3335 | 0.3723 | 0.0847 |
126
+ | 0.7408 | 0.64 | 3200 | 0.2573 | 0.2840 | 0.0642 |
127
+ | 0.6516 | 0.8 | 4000 | 0.2365 | 0.2581 | 0.0595 |
128
+ | 0.6242 | 0.96 | 4800 | 0.2039 | 0.2433 | 0.0541 |
129
+ | 0.5754 | 1.12 | 5600 | 0.1832 | 0.2156 | 0.0482 |
130
+ | 0.5626 | 1.28 | 6400 | 0.1827 | 0.2091 | 0.0463 |
131
+ | 0.5342 | 1.44 | 7200 | 0.1744 | 0.2033 | 0.0468 |
132
+ | 0.4965 | 1.6 | 8000 | 0.1705 | 0.1963 | 0.0444 |
133
+ | 0.5047 | 1.76 | 8800 | 0.1604 | 0.1889 | 0.0422 |
134
+ | 0.4814 | 1.92 | 9600 | 0.1604 | 0.1827 | 0.0411 |
135
+ | 0.4471 | 2.09 | 10400 | 0.1566 | 0.1822 | 0.0406 |
136
+ | 0.4509 | 2.25 | 11200 | 0.1619 | 0.1853 | 0.0432 |
137
+ | 0.4415 | 2.41 | 12000 | 0.1513 | 0.1764 | 0.0397 |
138
+ | 0.4313 | 2.57 | 12800 | 0.1515 | 0.1739 | 0.0392 |
139
+ | 0.4163 | 2.73 | 13600 | 0.1445 | 0.1695 | 0.0377 |
140
+ | 0.4142 | 2.89 | 14400 | 0.1478 | 0.1699 | 0.0385 |
141
+ | 0.4184 | 3.05 | 15200 | 0.1430 | 0.1669 | 0.0376 |
142
+ | 0.3886 | 3.21 | 16000 | 0.1433 | 0.1644 | 0.0374 |
143
+ | 0.3795 | 3.37 | 16800 | 0.1426 | 0.1648 | 0.0373 |
144
+ | 0.3859 | 3.53 | 17600 | 0.1357 | 0.1604 | 0.0361 |
145
+ | 0.3762 | 3.69 | 18400 | 0.1344 | 0.1558 | 0.0349 |
146
+ | 0.384 | 3.85 | 19200 | 0.1379 | 0.1576 | 0.0359 |
147
+ | 0.3762 | 4.01 | 20000 | 0.1344 | 0.1539 | 0.0346 |
148
+ | 0.3559 | 4.17 | 20800 | 0.1339 | 0.1525 | 0.0351 |
149
+ | 0.3683 | 4.33 | 21600 | 0.1315 | 0.1518 | 0.0342 |
150
+ | 0.3572 | 4.49 | 22400 | 0.1307 | 0.1507 | 0.0342 |
151
+ | 0.3494 | 4.65 | 23200 | 0.1294 | 0.1491 | 0.0335 |
152
+ | 0.3476 | 4.81 | 24000 | 0.1287 | 0.1491 | 0.0336 |
153
+ | 0.3475 | 4.97 | 24800 | 0.1271 | 0.1475 | 0.0329 |
154
+
155
  ### Framework versions
156
 
157
  - Transformers 4.16.2