Whisper Small GA-EN Speech Translation + VAD + warmup_ratio=0.01
This model is a fine-tuned version of openai/whisper-small on the IWSLT-2023, FLEURS, BiteSize, and SpokenWords dataset. It achieves the following results on the evaluation set:
- Loss: 1.7482
- Bleu: 29.94
- Chrf: 45.74
- Wer: 64.3404
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: 0.0001
- train_batch_size: 64
- eval_batch_size: 64
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.01
- training_steps: 3000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Bleu | Chrf | Validation Loss | Wer |
|---|---|---|---|---|---|---|
| 2.0518 | 0.2188 | 100 | 8.56 | 25.29 | 1.8072 | 123.9982 |
| 1.5449 | 0.4376 | 200 | 18.41 | 34.82 | 1.5746 | 83.7461 |
| 1.2518 | 0.6565 | 300 | 21.1 | 36.24 | 1.5009 | 83.9712 |
| 1.0947 | 0.8753 | 400 | 21.5 | 41.43 | 1.4582 | 89.8694 |
| 0.4439 | 1.0941 | 500 | 25.21 | 41.77 | 1.4979 | 72.5799 |
| 0.4416 | 1.3129 | 600 | 22.2 | 40.47 | 1.5107 | 79.8739 |
| 0.4417 | 1.5317 | 700 | 20.2 | 40.75 | 1.5215 | 88.8789 |
| 0.4108 | 1.7505 | 800 | 25.73 | 41.28 | 1.5278 | 67.8073 |
| 0.355 | 1.9694 | 900 | 20.6 | 39.37 | 1.5436 | 87.3030 |
| 0.1303 | 2.1882 | 1000 | 28.79 | 42.68 | 1.5936 | 68.1675 |
| 0.1421 | 2.4070 | 1100 | 27.84 | 42.58 | 1.5745 | 67.5371 |
| 0.1341 | 2.6258 | 1200 | 30.52 | 45.15 | 1.5953 | 66.5916 |
| 0.1365 | 2.8446 | 1300 | 26.93 | 43.72 | 1.6046 | 74.2909 |
| 0.0528 | 3.0635 | 1400 | 29.03 | 44.12 | 1.6303 | 64.8807 |
| 0.0519 | 3.2823 | 1500 | 27.75 | 44.34 | 1.6774 | 68.6177 |
| 0.0554 | 3.5011 | 1600 | 27.64 | 45.15 | 1.6637 | 71.1842 |
| 0.0514 | 3.7199 | 1700 | 30.26 | 44.62 | 1.6497 | 65.4660 |
| 0.0503 | 3.9387 | 1800 | 26.88 | 43.0 | 1.6780 | 70.4187 |
| 0.0259 | 4.1575 | 1900 | 29.6 | 44.51 | 1.6915 | 64.9707 |
| 0.0263 | 4.3764 | 2000 | 25.33 | 42.51 | 1.7080 | 72.3998 |
| 0.0254 | 4.5952 | 2100 | 30.59 | 45.35 | 1.6884 | 64.2954 |
| 0.0211 | 4.8140 | 2200 | 31.09 | 46.56 | 1.6984 | 64.0252 |
| 0.0137 | 5.0328 | 2300 | 28.96 | 43.67 | 1.7253 | 66.3665 |
| 0.0075 | 5.2516 | 2400 | 29.77 | 44.63 | 1.7112 | 66.9968 |
| 0.0056 | 5.4705 | 2500 | 29.96 | 45.51 | 1.7197 | 64.5655 |
| 0.0067 | 5.6893 | 2600 | 29.86 | 45.25 | 1.7464 | 66.0964 |
| 0.0064 | 5.9081 | 2700 | 29.47 | 45.36 | 1.7440 | 65.2859 |
| 0.0023 | 6.1269 | 2800 | 30.03 | 46.49 | 1.7419 | 64.4755 |
| 0.0016 | 6.3457 | 2900 | 29.76 | 45.64 | 1.7474 | 65.0158 |
| 0.0019 | 6.5646 | 3000 | 1.7482 | 29.94 | 45.74 | 64.3404 |
Framework versions
- Transformers 4.41.2
- Pytorch 2.2.0+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1