wav2vec2-large-xlsr-53-german-cv9
This model is a fine-tuned version of ./facebook/wav2vec2-large-xlsr-53 on the MOZILLA-FOUNDATION/COMMON_VOICE_9_0 - DE dataset.
It achieves the following results on the test set:
- CER: 2.273015898213336
- Wer: 9.480663281840769
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 16
- eval_batch_size: 32
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 50.0
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Eval Wer |
---|---|---|---|---|
0.4129 | 1.0 | 3557 | 0.3015 | 0.2499 |
0.2121 | 2.0 | 7114 | 0.1596 | 0.1567 |
0.1455 | 3.0 | 10671 | 0.1377 | 0.1354 |
0.1436 | 4.0 | 14228 | 0.1301 | 0.1282 |
0.1144 | 5.0 | 17785 | 0.1225 | 0.1245 |
0.1219 | 6.0 | 21342 | 0.1254 | 0.1208 |
0.104 | 7.0 | 24899 | 0.1198 | 0.1232 |
0.1016 | 8.0 | 28456 | 0.1149 | 0.1174 |
0.1093 | 9.0 | 32013 | 0.1186 | 0.1186 |
0.0858 | 10.0 | 35570 | 0.1182 | 0.1164 |
0.102 | 11.0 | 39127 | 0.1191 | 0.1186 |
0.0834 | 12.0 | 42684 | 0.1161 | 0.1096 |
0.0916 | 13.0 | 46241 | 0.1147 | 0.1107 |
0.0811 | 14.0 | 49798 | 0.1174 | 0.1136 |
0.0814 | 15.0 | 53355 | 0.1132 | 0.1114 |
0.0865 | 16.0 | 56912 | 0.1134 | 0.1097 |
0.0701 | 17.0 | 60469 | 0.1096 | 0.1054 |
0.0891 | 18.0 | 64026 | 0.1110 | 0.1076 |
0.071 | 19.0 | 67583 | 0.1141 | 0.1074 |
0.0726 | 20.0 | 71140 | 0.1094 | 0.1093 |
0.0647 | 21.0 | 74697 | 0.1088 | 0.1095 |
0.0643 | 22.0 | 78254 | 0.1105 | 0.1044 |
0.0764 | 23.0 | 81811 | 0.1072 | 0.1042 |
0.0605 | 24.0 | 85368 | 0.1095 | 0.1026 |
0.0722 | 25.0 | 88925 | 0.1144 | 0.1066 |
0.0597 | 26.0 | 92482 | 0.1087 | 0.1022 |
0.062 | 27.0 | 96039 | 0.1073 | 0.1027 |
0.0536 | 28.0 | 99596 | 0.1068 | 0.1027 |
0.0616 | 29.0 | 103153 | 0.1097 | 0.1037 |
0.0642 | 30.0 | 106710 | 0.1117 | 0.1020 |
0.0555 | 31.0 | 110267 | 0.1109 | 0.0990 |
0.0632 | 32.0 | 113824 | 0.1104 | 0.0977 |
0.0482 | 33.0 | 117381 | 0.1108 | 0.0958 |
0.0601 | 34.0 | 120938 | 0.1095 | 0.0957 |
0.0508 | 35.0 | 124495 | 0.1079 | 0.0973 |
0.0526 | 36.0 | 128052 | 0.1068 | 0.0967 |
0.0487 | 37.0 | 131609 | 0.1081 | 0.0966 |
0.0495 | 38.0 | 135166 | 0.1099 | 0.0956 |
0.0528 | 39.0 | 138723 | 0.1091 | 0.0923 |
0.0439 | 40.0 | 142280 | 0.1111 | 0.0928 |
0.0467 | 41.0 | 145837 | 0.1131 | 0.0943 |
0.0407 | 42.0 | 149394 | 0.1115 | 0.0944 |
0.046 | 43.0 | 152951 | 0.1106 | 0.0935 |
0.0447 | 44.0 | 156508 | 0.1083 | 0.0919 |
0.0434 | 45.0 | 160065 | 0.1093 | 0.0909 |
0.0472 | 46.0 | 163622 | 0.1092 | 0.0921 |
0.0414 | 47.0 | 167179 | 0.1106 | 0.0922 |
0.0501 | 48.0 | 170736 | 0.1094 | 0.0918 |
0.0388 | 49.0 | 174293 | 0.1099 | 0.0918 |
0.0428 | 50.0 | 177850 | 0.1103 | 0.0915 |
Framework versions
- Transformers 4.19.0.dev0
- Pytorch 1.11.0+cu113
- Datasets 2.0.0
- Tokenizers 0.11.6
- Downloads last month
- 110
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.
Dataset used to train oliverguhr/wav2vec2-large-xlsr-53-german-cv9
Evaluation results
- Test WER on Common Voice 9self-reported9.481
- Test CER on Common Voice 9self-reported1.917
- Test WER (+LM) on Common Voice 9self-reported7.490
- Test CER (+LM) on Common Voice 9self-reported1.917
- Test WER on Common Voice 6.1self-reported8.122
- Test CER on Common Voice 6.1self-reported1.000
- Test WER (+LM) on Common Voice 6.1self-reported6.145
- Test CER (+LM) on Common Voice 6.1self-reported1.525