NoLoR: An ASR-Based Framework for Expedited Endangered Language Documentation with Neo-Aramaic as a Case Study

What it is about

NoLoR is a four-step loop for documenting an unwritten or under-documented language: define a phonemic orthography, build a small initial dataset of transcribed speech, fine-tune a pretrained speech model, and then expand the dataset with the model's drafts, which people correct. The case is the Neo-Aramaic dialect of Urmi, with a 35-minute labelled set cut into about 600 clips of at most 15 seconds, and a wav2vec 2.0 checkpoint first fine-tuned on Persian. The model reaches a character error rate of 12.5 per cent on the test data. With the draft, transcription is 2.0 times faster for 15-second clips and up to 6.3 times faster for 5-minute clips. A subword tokenizer with a small vocabulary did worse than characters on so little text, and the single-speaker data was augmented for speed, pitch and microphone clarity.

Ideas

• I1. Time the correction of a parser draft against writing the tree from scratch, on items of growing length, and report the speed-up in the same form as the paper's table. For Mova the draft comes from gparse over treebanks/hbo/ptnk, and the timing is run twice to check it is stable.
• I2. Two spellings for one language: a phonemic working orthography for the data, and a simplified keyboard spelling for readers (the paper writes a sh sound as the English digraph). Use the same split for an Aramaic variety, with the phonemic form as the master and the simple form as a display layer.
• I3. Small data favours tables over learned subwords. The paper found that a 256 or 512 subword vocabulary failed on a little text. For the tools/hbo segmenter, which uses learned prefix and suffix tables, compare it against a subword split on the held-out PTNK sentences and keep whichever is better on the small slice.

For Mova

I1 goes into treebanks/hbo/ptnk as a correction-time measure for gparse drafts. I2 goes into Script and phonology, with the simplified layer noted for languages/jpa/. I3 goes into tools/hbo, in the segmenter comparison.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. arXiv 2412.04717

@misc{nazari2024nolor,
  title = {{NoLoR: An ASR-Based Framework for Expedited Endangered Language Documentation with Neo-Aramaic as a Case Study}},
  author = {Nazari, Matthew},
  year = {2024},
  eprint = {2412.04717},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.04717}
}

References

26 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Adams, O.; Makarucha, A.; Neubig, G.; Bird, S.; and Cohn, T. 2017. Cross-Lingual Word Embeddings for Low-Resource Language Modeling. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, 937– 947. Valencia, Spain: Association for Computational Linguistics. ↗
2. Ardila, R.; Branson, M.; Davis, K.; Henretty, M.; Kohler, M.; Meyer, J.; Morais, R.; Saunders, L.; Tyers, F. M.; and Weber, G. 2019. Common Voice: A Massively-Multilingual Speech Corpus. arXiv. find
3. Baevski, A.; Zhou, H.; Mohamed, A.; and Auli, M. 2020. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. Bosch, L. t. 2000. Emotions: what is possible in the ASR framework. In ISCA Tutorial and Research Workshop (ITRW) on Speech and Emotion. find
4. Cahill, M.; and Karan, E. 2008. Factors in designing effective orthographies for unwritten languages. SIL International. find
5. Chen, L.-W.; and Rudnicky, A. 2021. Exploring Wav2vec 2.0 fine-tuning for improved speech emotion recognition. arXiv preprint arXiv:2110.06309. ↗
6. Feng, S.; Żelasko, P.; Moro-Velázquez, L.; Abavisani, A.; Hasegawa-Johnson, M.; Scharenborg, O.; and Dehak, N. 2021a. How Phonotactics Affect Multilingual and Zero-Shot ASR Performance. In ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 7238–7242. ↗
7. Feng, S. Y.; Gangal, V.; Wei, J.; Chandar, S.; Vosoughi, S.; Mitamura, T.; and Hovy, E. 2021b. A Survey of Data Augmentation Approaches for NLP. ↗
8. Gaunt, D. 2020. The Long Assyrian Genocide. Collective and State Violence in Turkey, 56–96. ↗
9. Graves, A. 2012. Connectionist Temporal Classification. Berlin, Heidelberg: Springer Berlin Heidelberg. ISBN 978-3-642-24797-2. find
10. Hedderich, M. A.; Lange, L.; Adel, H.; Strötgen, J.; and Klakow, D. 2020. A Survey on Recent Approaches for Natural Language Processing in Low-Resource Scenarios. ↗
11. Khan, G. 2016. The Neo-Aramaic Dialect of the Assyrian Christians of Urmi (4 vols). Brill. ↗
12. Khare, S.; Mittal, A. R.; Diwan, A.; Sarawagi, S.; Jyothi, P.; and Bharadwaj, S. 2021. Low Resource ASR: The Surprising Effectiveness of High Resource Transliteration. In Interspeech, 1529–1533. ↗
13. Malik, M.; Malik, M. K.; Mehmood, K.; and Makhdoom, I. 2021. Automatic speech recognition: a survey. Multimedia Tools and Applications, 80(6): 9411–9457. ↗
14. Moseley, C.; and Nicolas, A. 2010. Atlas of the world’s languages in Danger. UNESCO Pub. find
15. Neyshabur, B.; Bhojanapalli, S.; McAllester, D.; and Srebro, N. 2017. Exploring generalization in deep learning. Advances in neural information processing systems, 30. find
16. NLLB Team; Costa-jussà, M. R.; Cross, J.; Çelebi, O.; Elbayad, M.; Heafield, K.; Heffernan, K.; Kalbassi, E.; Lam, J.; Licht, D.; Maillard, J.; Sun, A.; Wang, S.; Wenzek, G.; Youngblood, A.; Akula, B.; Barrault, L.; Gonzalez, G. M.; Hansanti, P.; Hoffman, J.; Jarrett, S.; Sadagopan, K. R.; Rowe, D.; Spruit, S.; Tran, C.; Andrews, P.; Ayan, N. F.; Bhosale, S.; Edunov, S.; Fan, A.; Gao, C.; Goswami, V.; Guzmán, F.; Koehn, P.; Mourachko, A.; Ropers, C.; Saleem, S.; Schwenk, H.; and Wang, J. 2022. No Language Left Behind: Scaling Human-Centered Machine Translation. find
17. Omer, M. 2022. Religious Freedom in Iraq in 2021. find
18. Pike, K. L. 1971. Phonemics: A technique for reducing languages to writing. find
19. Ranathunga, S.; Lee, E.-S. A.; Skenduli, M. P.; Shekhar, R.; Alam, M.; and Kaur, R. 2021. Neural Machine Translation for Low-Resource Languages: A Survey. find
20. Shi, J.; Amith, J. D.; Garcı́a, R. C.; Sierra, E. G.; Duh, K.; and Watanabe, S. 2021. Leveraging End-to-End ASR for Endangered Language Documentation: An Empirical Study on Yoloxóchitl Mixtec. ↗
21. SIL International. 2007. FLEx. find
22. SIL International. 2010. SayMore. find
23. Venezky, R. L. 2004. In search of the perfect orthography. Written Language & Literacy, 7(2): 139–163. ↗
24. Vicsi, K.; and Szaszák, G. 2006. Prosodic cues for automatic phrase boundary detection in ASR. In International Conference on Text, Speech and Dialogue, 547–554. Springer. ↗
25. Xu, J.; Tan, X.; Ren, Y.; Qin, T.; Li, J.; Zhao, S.; and Liu, T.-Y. 2020. LRSpeech: Extremely Low-Resource Speech Synthesis and Recognition. ↗
26. Zahrer, A.; Zgank, A.; and Schuppler, B. 2020. Towards Building an Automatic Transcription System for Language Documentation: Experiences from Muyu. In Proceedings of the Twelfth Language Resources and Evaluation Conference, 2893–2900. Marseille, France: European Language Resources Association. ISBN 979-10-95546-34-4. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/langstream-2024-nazari-nolor-asr-based-framework-expedited/