|
--- |
|
datasets: |
|
- Bingsu/zeroth-korean |
|
language: |
|
- ko |
|
metrics: |
|
- wer |
|
--- |
|
# Whisper Small Kr |
|
|
|
This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) on the Bingsu korean dataset. It achieves the following results: |
|
|
|
### Training results |
|
| Step | Training Loss | Validation Loss | Wer | |
|
|------|---------------|-----------------|-----------| |
|
| 1000 | 0.1141 | 0.145484 | 10.811625 | |
|
| 2000 | 0.0369 | 0.106371 | 7.72474 | |
|
| 3000 | 0.0153 | 0.095729 | 6.520102 | |
|
|
|
### 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: 3000 |
|
- mixed_precision_training: Native AMP |
