Description
This model is a fine-tuned version of openai/whisper-small on Bangla speech data.
Environment:
- Python version: 3.12.12
- PyTorch version: 2.8.0+cu126
- NumPy version: 1.26.4
Training Parameters:
- Total epochs: 8
- Learning Rate: 2e-5
- Batch size per device: 4
- Gradient accumulation: 4
- Warmup Steps: 200
- LR Scheduler: cosine_with_restarts
Validation Set Evaluation:
| Epoch | Training Loss | Validation Loss | WER (%) | Normalized Levenshtein Similarity (%) |
|---|---|---|---|---|
| 0 | 1.436900 | 1.500822 | 23.480663 | 84.353535 |
| 2 | 1.433300 | 1.479616 | 20.883978 | 85.525253 |
| 4 | 1.418600 | 1.467377 | 16.022099 | 88.111111 |
| 6 | 1.413600 | 1.469696 | 14.751381 | 89.313131 |
| 8 | 1.413300 | 1.470008 | 14.530387 | 89.444444 |