Whisper Medium GA-EN Speech Translation
This model is a fine-tuned version of openai/whisper-medium on the IWSLT-2023, FLEURS, BiteSize, SpokenWords, Tatoeba, Wikimedia, and EUbookshop dataset. It achieves the following results on the evaluation set:
- Loss: 1.0552
- Bleu: 33.24
- Chrf: 55.16
- Wer: 61.5038
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: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.03
- training_steps: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Bleu | Chrf | Wer |
|---|---|---|---|---|---|---|
| 2.5219 | 0.0138 | 100 | 2.1106 | 0.44 | 10.48 | 107.2490 |
| 2.4608 | 0.0276 | 200 | 2.1816 | 3.3 | 20.43 | 179.1535 |
| 2.3008 | 0.0414 | 300 | 2.0587 | 3.66 | 21.59 | 206.4836 |
| 2.2095 | 0.0552 | 400 | 1.9459 | 8.79 | 27.66 | 100.3602 |
| 2.0454 | 0.0690 | 500 | 1.8681 | 8.14 | 27.36 | 122.1522 |
| 1.9937 | 0.0828 | 600 | 1.8717 | 11.05 | 30.26 | 97.2535 |
| 1.868 | 0.0966 | 700 | 1.7917 | 9.14 | 29.03 | 129.0410 |
| 1.9924 | 0.1103 | 800 | 1.7170 | 12.62 | 33.2 | 89.6443 |
| 1.8646 | 0.1241 | 900 | 1.7252 | 11.98 | 30.77 | 97.8838 |
| 1.7644 | 0.1379 | 1000 | 1.6832 | 10.87 | 31.0 | 109.1851 |
| 1.692 | 0.1517 | 1100 | 1.6837 | 13.05 | 34.46 | 93.3814 |
| 1.7044 | 0.1655 | 1200 | 1.5527 | 20.95 | 37.42 | 75.2364 |
| 1.6824 | 0.1793 | 1300 | 1.5611 | 14.91 | 35.56 | 92.6159 |
| 1.6557 | 0.1931 | 1400 | 1.5554 | 14.0 | 36.54 | 99.8199 |
| 1.5456 | 0.2069 | 1500 | 1.5058 | 19.72 | 39.81 | 83.5660 |
| 1.3755 | 0.2207 | 1600 | 1.5039 | 18.04 | 37.95 | 82.9806 |
| 1.3959 | 0.2345 | 1700 | 1.4374 | 17.01 | 39.5 | 85.2319 |
| 1.5012 | 0.2483 | 1800 | 1.4242 | 14.93 | 39.24 | 114.4079 |
| 1.4278 | 0.2621 | 1900 | 1.3904 | 23.85 | 42.69 | 73.0302 |
| 1.3285 | 0.2759 | 2000 | 1.4493 | 17.7 | 37.23 | 83.8811 |
| 1.2655 | 0.2897 | 2100 | 1.3661 | 20.1 | 40.32 | 79.7839 |
| 1.2074 | 0.3034 | 2200 | 1.3387 | 24.45 | 43.79 | 72.9851 |
| 1.1893 | 0.3172 | 2300 | 1.3308 | 21.45 | 42.61 | 82.3953 |
| 1.1236 | 0.3310 | 2400 | 1.3050 | 22.77 | 44.17 | 77.3075 |
| 1.0934 | 0.3448 | 2500 | 1.2793 | 25.54 | 46.32 | 72.2647 |
| 1.06 | 0.3586 | 2600 | 1.2396 | 28.27 | 47.32 | 65.6911 |
| 1.0327 | 0.3724 | 2700 | 1.2577 | 28.45 | 47.01 | 67.3570 |
| 1.1623 | 0.3862 | 2800 | 1.2194 | 24.54 | 47.43 | 73.6155 |
| 1.0215 | 0.4 | 2900 | 1.2039 | 27.4 | 49.6 | 69.2481 |
| 0.9185 | 0.4138 | 3000 | 1.1724 | 27.04 | 49.24 | 67.8973 |
| 0.9003 | 0.4276 | 3100 | 1.1674 | 31.08 | 50.11 | 63.8001 |
| 0.9839 | 0.4414 | 3200 | 1.1580 | 30.24 | 50.63 | 64.5655 |
| 0.9396 | 0.4552 | 3300 | 1.1202 | 30.79 | 51.72 | 64.9257 |
| 0.9051 | 0.4690 | 3400 | 1.1180 | 30.34 | 53.08 | 66.4566 |
| 0.8621 | 0.4828 | 3500 | 1.1042 | 33.3 | 53.86 | 60.7834 |
| 0.8236 | 0.4966 | 3600 | 1.1070 | 32.77 | 53.21 | 62.0441 |
| 0.829 | 0.5103 | 3700 | 1.0771 | 32.49 | 54.21 | 62.5844 |
| 0.8375 | 0.5241 | 3800 | 1.0780 | 32.27 | 53.98 | 63.0797 |
| 0.8206 | 0.5379 | 3900 | 1.0615 | 33.26 | 55.07 | 61.6389 |
| 0.8059 | 0.5517 | 4000 | 1.0552 | 33.24 | 55.16 | 61.5038 |
Framework versions
- Transformers 4.41.2
- Pytorch 2.2.0+cu121
- Datasets 2.20.0
- Tokenizers 0.19.1
Citation
@inproceedings{moslem-2024-leveraging,
title = "Leveraging Synthetic Audio Data for End-to-End Low-Resource Speech Translation",
author = "Moslem, Yasmin",
booktitle = "Proceedings of the 21st International Conference on Spoken Language Translation (IWSLT 2024)",
month = aug,
year = "2024",
address = "Bangkok, Thailand (in-person and online)",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.iwslt-1.31/",
doi = "10.18653/v1/2024.iwslt-1.31",
pages = "265--273",
abstract = "This paper describes our system submission to the International Conference on Spoken Language Translation (IWSLT 2024) for Irish-to-English speech translation. We built end-to-end systems based on Whisper, and employed a number of data augmentation techniques, such as speech back-translation and noise augmentation. We investigate the effect of using synthetic audio data and discuss several methods for enriching signal diversity."
}