Model Details
Bangla dialect ASR (Automatic Speech Recognition) Model
Training Procedure
Bangla ASR Model trained for the contest https://www.kaggle.com/competitions/shobdotori The original dataset was augmented and then trained with an extra dataset
Model Sources:
Training Hyperparameters
- Batch Size: 4
- Learning rate: 2e-5
- Warmup Steps: 200
- Epochs: 8
- Gradient Accumulation Steps: 4
- LR Scheduler: cosine_with_restarts
Evaluation
| Epoch | Training Loss | Validation Loss | WER | Normalized Levenshtein Similarity |
|---|---|---|---|---|
| 0 | 1.480600 | 1.456514 | 6.940371 | 94.997286 |
| 2 | 1.445800 | 1.432858 | 4.838710 | 96.562330 |
| 4 | 1.428500 | 1.428582 | 2.932551 | 97.937398 |
| 6 | 1.422000 | 1.425159 | 2.981427 | 97.792654 |
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