A Quick-trained Whisper-Small model for Nan-TW (閩南話/台語) #JL
This model is a fine-tuned version of openai/whisper-small on the Common Voice 17.0 (nan-tw) dataset. It achieves the following results on the evaluation set:
- Loss: 0.7699
- Cer: 138.9186
Transcription Example
(Example Source: https://sutian.moe.edu.tw/zh-hant/su/27169/)
Original sentence: 萬事起頭難。
Inference by Whisper-Small: บันซู ขี้เท่าหลัน
Inference by this model: 萬事起頭難
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: 8
- seed: 42
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Cer |
|---|---|---|---|---|
| 0.3511 | 2.9240 | 1000 | 0.7512 | 125.6361 |
| 0.0117 | 5.8480 | 2000 | 0.7479 | 141.2850 |
| 0.001 | 8.7719 | 3000 | 0.7629 | 136.0814 |
| 0.0006 | 11.6959 | 4000 | 0.7699 | 138.9186 |
Framework versions
- Transformers 4.51.3
- Pytorch 2.7.0+cu126
- Datasets 3.5.1
- Tokenizers 0.21.1