Fine tuned Model can be found here: On huggingface
This Model is a fine-tuned version of Ranjit/Whisper_Small_Odia_10k_steps on the mozilla-foundation/common_voice_11_0 or dataset. It achieves the following results on the evaluation set:
- Loss: 0.4827
- Wer: 23.4979
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-06
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 200
- training_steps: 5000
- mixed_precision_training: Native AMP
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.0018 | 50.0 | 1000 | 0.3315 | 24.0903 |
| 0.0 | 100.0 | 2000 | 0.4098 | 23.7236 |
| 0.0 | 150.0 | 3000 | 0.4827 | 23.4979 |
| 0.0 | 200.0 | 4000 | 0.4914 | 23.8928 |
| 0.0 | 250.0 | 5000 | 0.4953 | 23.7800 |