This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the librispeech_asr dataset. It achieves the following results on the evaluation set:
- Loss: 0.1444
- Wer: 0.1167
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: 5e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 32
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 1000
- num_epochs: 50
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Wer |
---|---|---|---|---|
2.9365 | 4.17 | 500 | 2.9398 | 0.9999 |
1.5444 | 8.33 | 1000 | 0.5947 | 0.4289 |
1.1367 | 12.5 | 1500 | 0.2751 | 0.2366 |
0.9972 | 16.66 | 2000 | 0.2032 | 0.1797 |
0.9118 | 20.83 | 2500 | 0.1786 | 0.1479 |
0.8664 | 24.99 | 3000 | 0.1641 | 0.1408 |
0.8251 | 29.17 | 3500 | 0.1537 | 0.1267 |
0.793 | 33.33 | 4000 | 0.1525 | 0.1244 |
0.785 | 37.5 | 4500 | 0.1470 | 0.1184 |
0.7612 | 41.66 | 5000 | 0.1446 | 0.1177 |
0.7478 | 45.83 | 5500 | 0.1449 | 0.1176 |
0.7443 | 49.99 | 6000 | 0.1444 | 0.1167 |
Framework versions
- Transformers 4.17.0.dev0
- Pytorch 1.10.2+cu102
- Datasets 1.18.2.dev0
- Tokenizers 0.11.0
- Downloads last month
- 27
Inference Providers
NEW
This model is not currently available via any of the supported third-party Inference Providers, and
the model is not deployed on the HF Inference API.
Model tree for vitouphy/wav2vec2-xls-r-300m-english
Dataset used to train vitouphy/wav2vec2-xls-r-300m-english
Evaluation results
- Test WER on LibriSpeech (clean)test set self-reported12.290
- Test CER on LibriSpeech (clean)test set self-reported3.340
- Validation WER on Robust Speech Event - Dev Dataself-reported36.750
- Validation CER on Robust Speech Event - Dev Dataself-reported14.830
- Test WER on Common Voice 8.0test set self-reported37.810
- Test WER on Robust Speech Event - Test Dataself-reported38.800