wav2vec2-xls-r-300m-Br-small
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset. It achieves the following results on the evaluation set:
- Loss: 1.0573
- Wer: 0.6675
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: 0.0003
- train_batch_size: 16
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- 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: 500
- num_epochs: 30
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 5.7464 | 2.79 | 400 | 1.7474 | 1.1018 |
| 1.1117 | 5.59 | 800 | 0.9434 | 0.8697 |
| 0.6481 | 8.39 | 1200 | 0.9251 | 0.7910 |
| 0.4754 | 11.19 | 1600 | 0.9208 | 0.7412 |
| 0.3602 | 13.98 | 2000 | 0.9284 | 0.7232 |
| 0.2873 | 16.78 | 2400 | 0.9299 | 0.6940 |
| 0.2386 | 19.58 | 2800 | 1.0182 | 0.6927 |
| 0.1971 | 22.38 | 3200 | 1.0456 | 0.6898 |
| 0.1749 | 25.17 | 3600 | 1.0208 | 0.6769 |
| 0.1487 | 27.97 | 4000 | 1.0573 | 0.6675 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu111
- Datasets 1.14.0
- Tokenizers 0.10.3
