Whisper Large GA-EN Speech Translation
This model is a fine-tuned version of openai/whisper-large on the IWSLT-2023, FLEURS, BiteSize, SpokenWords, Tatoeba, and Wikimedia dataset. The datasets are augmented in two ways: noise augmentation, and truncating low-amplitude samples. The best model checkpoint (this version) based on ChrF is at step 3000, epoch 0.99, and it achieves the following results on the evaluation set:
- Loss: 1.1742
- Bleu: 30.16
- Chrf: 50.72
- Wer: 69.9685
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: 4
- eval_batch_size: 4
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 0.03
- training_steps: 3000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Bleu | Chrf | Wer |
|---|---|---|---|---|---|---|
| 3.1833 | 0.03 | 100 | 2.5169 | 2.03 | 16.8 | 215.5786 |
| 2.7632 | 0.07 | 200 | 2.1827 | 7.81 | 24.07 | 113.1022 |
| 2.5687 | 0.1 | 300 | 2.0746 | 6.16 | 24.2 | 158.8474 |
| 2.5615 | 0.13 | 400 | 1.9379 | 8.68 | 26.18 | 120.8465 |
| 2.4554 | 0.16 | 500 | 1.8932 | 12.14 | 28.94 | 103.1067 |
| 2.3546 | 0.2 | 600 | 1.8734 | 14.34 | 29.83 | 91.5353 |
| 2.2804 | 0.23 | 700 | 1.8075 | 13.18 | 33.07 | 105.5380 |
| 2.1408 | 0.26 | 800 | 1.7034 | 13.01 | 33.0 | 89.4642 |
| 2.0411 | 0.3 | 900 | 1.6556 | 16.73 | 34.97 | 91.4453 |
| 1.7766 | 0.33 | 1000 | 1.6505 | 17.21 | 35.54 | 83.5209 |
| 1.7704 | 0.36 | 1100 | 1.5800 | 17.54 | 38.11 | 77.1724 |
| 1.6537 | 0.39 | 1200 | 1.5684 | 14.2 | 35.39 | 95.6326 |
| 1.4841 | 0.43 | 1300 | 1.4970 | 22.96 | 39.35 | 71.3643 |
| 1.641 | 0.46 | 1400 | 1.4693 | 16.3 | 37.69 | 103.7821 |
| 1.393 | 0.49 | 1500 | 1.3923 | 27.21 | 43.87 | 69.3381 |
| 1.249 | 0.53 | 1600 | 1.3876 | 23.33 | 42.26 | 76.5421 |
| 1.3385 | 0.56 | 1700 | 1.3404 | 23.86 | 42.82 | 75.0563 |
| 1.2537 | 0.59 | 1800 | 1.3226 | 17.03 | 41.72 | 100.1801 |
| 1.2891 | 0.62 | 1900 | 1.2995 | 27.26 | 43.62 | 69.1580 |
| 1.226 | 0.66 | 2000 | 1.2605 | 30.89 | 47.34 | 63.5750 |
| 1.1268 | 0.69 | 2100 | 1.2783 | 27.43 | 45.97 | 67.4921 |
| 1.0007 | 0.72 | 2200 | 1.2521 | 27.21 | 47.25 | 71.0041 |
| 0.9565 | 0.76 | 2300 | 1.2219 | 31.65 | 48.07 | 64.2053 |
| 0.9309 | 0.79 | 2400 | 1.2193 | 31.4 | 48.18 | 64.1603 |
| 0.7923 | 0.82 | 2500 | 1.2099 | 28.88 | 48.89 | 69.7884 |
| 0.8199 | 0.85 | 2600 | 1.1972 | 29.37 | 48.07 | 67.3120 |
| 0.6974 | 0.89 | 2700 | 1.1857 | 29.7 | 48.95 | 70.5988 |
| 0.6736 | 0.92 | 2800 | 1.1884 | 29.33 | 48.97 | 72.7150 |
| 0.6826 | 0.95 | 2900 | 1.1834 | 30.76 | 50.11 | 68.1225 |
| 0.7001 | 0.99 | 3000 | 1.1742 | 30.16 | 50.72 | 69.9685 |
Framework versions
- Transformers 4.39.3
- Pytorch 2.0.1+cu118
- Datasets 2.18.0
- Tokenizers 0.15.2