<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->
Whisper Small Seneca
This model is a fine-tuned version of openai/whisper-medium on the audiofolder dataset. It achieves the following results on the evaluation set:
- Loss: 0.5983
- Wer: 27.4983
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: 2
- seed: 42
- distributed_type: multi-GPU
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 7768
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.1151 | 3.01 | 1000 | 0.4013 | 31.8326 |
| 0.0185 | 6.02 | 2000 | 0.4796 | 30.1660 |
| 0.0062 | 9.03 | 3000 | 0.5143 | 30.0417 |
| 0.0022 | 12.04 | 4000 | 0.5469 | 28.6301 |
| 0.0004 | 16.01 | 5000 | 0.5695 | 27.9522 |
| 0.0001 | 19.01 | 6000 | 0.5891 | 27.5294 |
| 0.0002 | 22.02 | 7000 | 0.5983 | 27.4983 |
Framework versions
- Transformers 4.25.1
- Pytorch 1.13.1+cu117
- Datasets 2.8.0
- Tokenizers 0.13.2
