wav2vec2-large-xls-r-300m-hungarian
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - HU dataset. It achieves the following results on the evaluation set:
- Loss: 0.2562
- Wer: 0.3112
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: 7e-05
- train_batch_size: 32
- eval_batch_size: 32
- 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: 50.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 2.3964 | 3.52 | 1000 | 1.2251 | 0.8781 |
| 1.3176 | 7.04 | 2000 | 0.3872 | 0.4462 |
| 1.1999 | 10.56 | 3000 | 0.3244 | 0.3922 |
| 1.1633 | 14.08 | 4000 | 0.3014 | 0.3704 |
| 1.1132 | 17.61 | 5000 | 0.2913 | 0.3623 |
| 1.0888 | 21.13 | 6000 | 0.2864 | 0.3498 |
| 1.0487 | 24.65 | 7000 | 0.2821 | 0.3435 |
| 1.0431 | 28.17 | 8000 | 0.2739 | 0.3308 |
| 0.9896 | 31.69 | 9000 | 0.2629 | 0.3243 |
| 0.9839 | 35.21 | 10000 | 0.2806 | 0.3308 |
| 0.9586 | 38.73 | 11000 | 0.2650 | 0.3235 |
| 0.9501 | 42.25 | 12000 | 0.2585 | 0.3173 |
| 0.938 | 45.77 | 13000 | 0.2561 | 0.3117 |
| 0.921 | 49.3 | 14000 | 0.2559 | 0.3115 |
Framework versions
- Transformers 4.16.0.dev0
- Pytorch 1.10.1+cu102
- Datasets 1.17.1.dev0
- Tokenizers 0.11.0