Whisper Small Danish - Robust
This model is a fine-tuned version of openai/whisper-small on the mozilla-foundation/common_voice_11_0 da dataset. It achieves the following results on the evaluation set:
- Loss: 0.7926
- Wer: 32.3251
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: 5e-05
- train_batch_size: 64
- 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: 500
- training_steps: 5000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.0232 | 15.15 | 1000 | 0.7538 | 35.5813 |
| 0.0061 | 30.3 | 2000 | 0.7933 | 34.3766 |
| 0.0016 | 45.45 | 3000 | 0.7993 | 33.5823 |
| 0.0003 | 60.61 | 4000 | 0.7986 | 31.6097 |
| 0.0002 | 75.76 | 5000 | 0.7901 | 32.1357 |
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.1+cu117
- Datasets 2.8.0
- Tokenizers 0.13.2