Model Details
Bangla dialect ASR (Automatic Speech Recognition) Model
Model Sources:
Training Hyperparameters
- Batch Size: 4
- Learning rate: 3e-5
- Warmup Steps: 750
- Epochs: 12
- Gradient Accumulation Steps: 4
- LR Scheduler: linear
- Weight decay: 0.005
Evaluation
| Epoch | Training Loss | Validation Loss | WER | Normalized Levenshtein Similarity |
|---|---|---|---|---|
| 0 | 1.079400 | 0.963404 | 39.897353 | 76.774691 |
| 2 | 0.816400 | 0.844433 | 23.716910 | 85.590027 |
| 4 | 0.777400 | 0.814212 | 17.504052 | 89.011984 |
| 6 | 0.756000 | 0.800245 | 16.072393 | 89.814135 |
| 8 | 0.746700 | 0.794167 | 12.884927 | 91.766032 |
| 10 | 0.742800 | 0.791846 | 11.696380 | 92.622563 |
| 11 | 0.742500 | 0.792748 | 11.939492 | 92.488058 |
Language and Environment
- Python version: 3.12.12
- PyTorch version: 2.8.0+cu126
- Transformers version: 4.44.2
- NumPy version: 1.26.4
CUDA Info
- GPU: Tesla T4
- GPU Memory: 15.83 GB