metadata
license: apache-2.0
tags:
- generated_from_trainer
metrics:
- wer
model-index:
- name: openai/whisper-medium
results: []
openai/whisper-medium
This model is a fine-tuned version of openai/whisper-medium on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.2029
- Wer: 8.3235
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: 2
- eval_batch_size: 1
- 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: 20000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.2652 | 0.1 | 2000 | 0.3469 | 15.3537 |
| 0.3273 | 0.2 | 4000 | 0.3151 | 14.1141 |
| 0.2696 | 0.3 | 6000 | 0.2955 | 13.2472 |
| 0.1725 | 0.4 | 8000 | 0.2787 | 11.6834 |
| 0.1741 | 0.5 | 10000 | 0.2648 | 11.0088 |
| 0.2037 | 0.6 | 12000 | 0.2470 | 10.1909 |
| 0.1586 | 0.7 | 14000 | 0.2333 | 9.4096 |
| 0.1548 | 0.8 | 16000 | 0.2184 | 8.9724 |
| 0.1799 | 1.08 | 18000 | 0.2064 | 8.2830 |
| 0.1165 | 1.18 | 20000 | 0.2029 | 8.3235 |
Framework versions
- Transformers 4.25.1
- Pytorch 1.10.0+cu102
- Datasets 2.7.1
- Tokenizers 0.13.2