openai/whisper-large-v2
This model is a fine-tuned version of openai/whisper-large-v2 on the None dataset. It achieves the following results on the evaluation set:
- Loss: 1.0925
- Wer: 41.4086
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: 8
- eval_batch_size: 4
- 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: 10000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.5216 | 1.04 | 1000 | 0.7054 | 58.7611 |
| 0.0872 | 3.02 | 2000 | 0.7803 | 60.1400 |
| 0.1073 | 4.06 | 3000 | 0.8312 | 61.0522 |
| 0.0617 | 6.04 | 4000 | 0.8583 | 48.2181 |
| 0.0053 | 8.02 | 5000 | 0.9135 | 41.8328 |
| 0.0049 | 9.06 | 6000 | 0.9697 | 43.3814 |
| 0.0044 | 11.04 | 7000 | 0.9863 | 41.9813 |
| 0.0006 | 13.02 | 8000 | 1.0359 | 42.7662 |
| 0.0019 | 14.06 | 9000 | 1.0714 | 41.3449 |
| 0.0007 | 16.04 | 10000 | 1.0925 | 41.4086 |
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.1+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2