Whisper Small GA-EN Speech Translation + VAD
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: 1.7352
- Bleu: 28.22
- Chrf: 44.19
- Wer: 68.5277
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
- training_steps: 3000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Bleu | Chrf | Wer |
|---|---|---|---|---|---|---|
| 1.9529 | 0.2188 | 100 | 1.7388 | 12.76 | 29.03 | 97.1184 |
| 1.5762 | 0.4376 | 200 | 1.5362 | 15.3 | 33.31 | 98.4241 |
| 1.2624 | 0.6565 | 300 | 1.4346 | 17.94 | 37.2 | 101.4408 |
| 1.0367 | 0.8753 | 400 | 1.4502 | 21.52 | 39.13 | 85.4120 |
| 0.4677 | 1.0941 | 500 | 1.4693 | 23.26 | 40.49 | 78.4331 |
| 0.4284 | 1.3129 | 600 | 1.5163 | 21.31 | 41.41 | 86.0873 |
| 0.4026 | 1.5317 | 700 | 1.4999 | 24.11 | 40.59 | 79.3787 |
| 0.4132 | 1.7505 | 800 | 1.5134 | 27.77 | 43.01 | 70.1936 |
| 0.3701 | 1.9694 | 900 | 1.5368 | 27.74 | 42.61 | 66.0964 |
| 0.1337 | 2.1882 | 1000 | 1.5692 | 27.96 | 43.77 | 64.9257 |
| 0.143 | 2.4070 | 1100 | 1.5516 | 26.06 | 42.12 | 71.3192 |
| 0.144 | 2.6258 | 1200 | 1.5839 | 27.55 | 43.19 | 69.7434 |
| 0.1372 | 2.8446 | 1300 | 1.5510 | 27.93 | 43.07 | 66.1414 |
| 0.0573 | 3.0635 | 1400 | 1.6567 | 26.34 | 41.69 | 72.3998 |
| 0.0554 | 3.2823 | 1500 | 1.6511 | 27.98 | 42.66 | 68.5277 |
| 0.0534 | 3.5011 | 1600 | 1.6732 | 28.29 | 43.2 | 67.1319 |
| 0.0588 | 3.7199 | 1700 | 1.6687 | 27.0 | 43.31 | 70.7789 |
| 0.0486 | 3.9387 | 1800 | 1.6759 | 28.02 | 43.97 | 66.3665 |
| 0.0224 | 4.1575 | 1900 | 1.7597 | 26.86 | 41.81 | 70.5538 |
| 0.0264 | 4.3764 | 2000 | 1.7113 | 27.58 | 43.38 | 70.4638 |
| 0.0233 | 4.5952 | 2100 | 1.7013 | 27.83 | 42.87 | 68.2575 |
| 0.0192 | 4.8140 | 2200 | 1.7351 | 25.39 | 42.09 | 78.0279 |
| 0.0149 | 5.0328 | 2300 | 1.7350 | 27.62 | 43.99 | 70.5538 |
| 0.0086 | 5.2516 | 2400 | 1.7331 | 29.37 | 45.08 | 68.5277 |
| 0.006 | 5.4705 | 2500 | 1.7145 | 29.04 | 44.19 | 66.9968 |
| 0.0064 | 5.6893 | 2600 | 1.7322 | 28.27 | 43.6 | 70.2386 |
| 0.0053 | 5.9081 | 2700 | 1.7239 | 27.86 | 43.78 | 69.6083 |
| 0.0021 | 6.1269 | 2800 | 1.7288 | 28.14 | 44.12 | 68.5727 |
| 0.0016 | 6.3457 | 2900 | 1.7375 | 28.26 | 44.14 | 68.7078 |
| 0.0023 | 6.5646 | 3000 | 1.7352 | 28.22 | 44.19 | 68.5277 |
Framework versions
- Transformers 4.41.1
- Pytorch 2.2.0+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1