ls-timit-wsj0-100percent-supervised-meta
This model was trained from scratch on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.0531
- Wer: 0.0214
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: 32
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 1000
- num_epochs: 20
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.1618 | 4.57 | 1000 | 0.0500 | 0.0432 |
| 0.0489 | 9.13 | 2000 | 0.0535 | 0.0291 |
| 0.0306 | 13.7 | 3000 | 0.0478 | 0.0275 |
| 0.0231 | 18.26 | 4000 | 0.0531 | 0.0214 |
Framework versions
- Transformers 4.16.2
- Pytorch 1.10.2
- Datasets 1.18.2
- Tokenizers 0.10.3