biodatlab/

distill-whisper-th-large-v3

Distilled Large-V3 Whisper ASR Model for Thai

Model Description

Fine-tuning Details

• Original Model: Whisper Large V3 Turbo
• Datasets Used for Fine-tuning:
  • Common Voice v13
  • Gowajee
  • Thai Elderly Speech Corpus
  • Custom Scraped Data
  • Thai-Central Dialect from SLSCU Thai Dialect Corpus

Model Performance

• DeepCut Tokenized WER on Common Voice 13 Test Set:
  • Original Model: 41.53%
  • This Model: 6.82%
• DeepCut Tokenized WER on FLEURS Test Set:
  • Original Model: 25.56%
  • This Model: 10.65%

Intended Use

Limitations

• The model is specifically trained for the Thai language and may not perform well with other languages.
• Performance might vary across different Thai dialects and accents.
• As with any ASR system, background noise and speech clarity can impact recognition accuracy.

Acknowledgments

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.2
• Datasets 2.16.1
• Tokenizers 0.15.0

Contributors

• Biomedical and Data Lab, Mahidol University
• WordSense by Looloo technology
• Mahamor by Looloo technology

Citation

@inproceedings{aung-etal-2024-thonburian,
    title = "Thonburian Whisper: Robust Fine-tuned and Distilled Whisper for {T}hai",
    author = "Aung, Zaw Htet  and
      Thavornmongkol, Thanachot  and
      Boribalburephan, Atirut  and
      Tangsriworakan, Vittavas  and
      Pipatsrisawat, Knot  and
      Achakulvisut, Titipat",
    editor = "Abbas, Mourad  and
      Freihat, Abed Alhakim",
    booktitle = "Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024)",
    month = oct,
    year = "2024",
    address = "Trento",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.icnlsp-1.17",
    pages = "149--156",
}