Whisper Tiny Tatar - Kirill Milintsevich
This model is a fine-tuned version of openai/whisper-tiny on the Common Voice 11.0 dataset. It achieves the following results on the evaluation set:
- Loss: 0.5106
- Wer: 49.2285
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: 64
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 5000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 0.4268 | 2.49 | 500 | 0.6232 | 63.6537 |
| 0.2331 | 4.98 | 1000 | 0.5044 | 52.3818 |
| 0.1332 | 7.46 | 1500 | 0.4927 | 50.2300 |
| 0.09 | 9.95 | 2000 | 0.5106 | 49.2285 |
| 0.048 | 12.44 | 2500 | 0.5526 | 49.7806 |
| 0.0346 | 14.93 | 3000 | 0.5850 | 50.0319 |
| 0.0181 | 17.41 | 3500 | 0.6276 | 50.5592 |
| 0.0122 | 19.9 | 4000 | 0.6494 | 50.3327 |
| 0.0086 | 22.39 | 4500 | 0.6737 | 50.6688 |
| 0.0077 | 24.88 | 5000 | 0.6777 | 50.6724 |