wav2vec2-300m-teste4
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: 0.3276
- Wer: 0.3489
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: 4
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 10.0237 | 0.49 | 100 | 4.2075 | 0.9792 |
| 3.313 | 0.98 | 200 | 3.0232 | 0.9792 |
| 2.9469 | 1.47 | 300 | 2.7591 | 0.9792 |
| 1.4217 | 1.96 | 400 | 0.8397 | 0.6219 |
| 0.5598 | 2.45 | 500 | 0.6085 | 0.5087 |
| 0.4507 | 2.94 | 600 | 0.4512 | 0.4317 |
| 0.2775 | 3.43 | 700 | 0.3839 | 0.3751 |
| 0.2047 | 3.92 | 800 | 0.3276 | 0.3489 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.9.1+cu102
- Datasets 1.17.0
- Tokenizers 0.10.3