Whisper Small TW on Chinese base
This model is a fine-tuned version of Jingmiao/whisper-small-chinese_base on the mozilla-foundation/common_voice_11_0 zh-TW dataset. It achieves the following results on the evaluation set:
- Loss: 0.2601
- Wer: 41.9038
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: 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 |
|---|---|---|---|---|
| 0.0071 | 6.02 | 1000 | 0.2364 | 42.6407 |
| 0.0008 | 13.02 | 2000 | 0.2601 | 41.9038 |
| 0.0004 | 20.01 | 3000 | 0.2771 | 42.3951 |
| 0.0003 | 27.0 | 4000 | 0.2867 | 42.6407 |
| 0.0002 | 33.02 | 5000 | 0.2901 | 42.6407 |
Framework versions
- Transformers 4.26.0.dev0
- Pytorch 1.13.1+cu117
- Datasets 2.8.1.dev0
- Tokenizers 0.13.2