Whisper Small Icelandic
This model is a fine-tuned version of openai/whisper-small on the samromur dataset. It achieves the following results on the evaluation set:
- Loss: 0.2613
- Wer: 23.0409
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: 1e-05
- train_batch_size: 16
- 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: 500
- training_steps: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.3551 | 0.18 | 1000 | 0.4322 | 35.0421 |
| 0.2541 | 0.36 | 2000 | 0.3249 | 27.4721 |
| 0.231 | 0.53 | 3000 | 0.2781 | 24.2234 |
| 0.2277 | 0.71 | 4000 | 0.2613 | 23.0409 |
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu116
- Datasets 2.7.1
- Tokenizers 0.13.2