Elamur Research⌕
☰
home / papers / 2025 / Exploring Transliteration-Based Zero-Shot Transfer for Amharic ASR

Exploring Transliteration-Based Zero-Shot Transfer for Amharic ASR

Paper · 2025 · Hellina Hailu Nigatu and Hanan Aldarmaki

Summary

Building speech recognition for low-resourced languages is hard because transcribed audio is scarce. Earlier zero-shot methods output IPA, which ordinary users cannot read, so they are a poor fit for captioning or dictation. The authors test a different route for Amharic. They convert the transcripts of other languages (Arabic, Xhosa, French, Spanish) into the Ge'ez script and train recognizers on that converted data. Xhosa and French/Spanish were transliterated with google-transliteration-api, and Arabic with a hand-built rule system. Two model families were trained: fine-tuned wav2vec2 XLS-R, and Kaldi-style GMM-HMM systems that decode with an Amharic lexicon and language model. With about 22 hours of source speech, the zero-shot systems beat previously reported Amharic results and produce native-script text directly. The best predictor of a good source language was not language family. It was how few test-set characters remained out-of-vocabulary after transliteration. When 10–60 minutes of Amharic data are added, transliterated pretraining clearly beats both untransliterated source data and Amharic-only fine-tuning. This makes it a practical augmentation strategy.

Key points

  • Amharic has 38 phonemes. Xhosa shares 30 of them, including the 5 ejectives, while French shares 23 and Spanish 21. Arabic was chosen for being in the same family.
  • The source data totals about 22 hours, roughly 4% of the data used in prior zero-shot work. Earlier Amharic zero-shot baselines include PER 74.8% (Feng et al.) and PTER 75.2% (Żelasko et al.), both on Babel.
  • GMM-HMM gives much lower WER in zero-shot mode because its decoder uses an Amharic lexicon. XLS-R yields lower CER.
  • With 10 minutes to 1 hour of Amharic data, transliterated source data cuts CER by up to ~30% absolute versus original-script source data. With 30 minutes of Amharic data, CER drops by ~40% and WER by ~10% absolute compared with zero-shot.
  • Limitations: the Arabic data comes from a single speaker and a narrow domain, because diacritized multi-speaker corpora were not available. Multilingual mixing was capped at two source languages, and only one target language was tested.

Interesting ideas

  • I1. Rank candidate source data by the share of target-side units it still leaves uncovered after mapping, not by genealogy; this cheap coverage check predicted transfer quality better than linguistic kinship.
  • I2. Approximate unmatched units softly by mapping to the nearest available target symbol, rather than discarding data that does not cover the full inventory.
  • I3. Train on source data while plugging in a target-language lexicon and language model only at decoding time, separating acoustic learning from target linguistic knowledge.
  • I4. Output directly in the user-facing representation (native script) instead of an intermediate notation that needs a later mapping step.
  • I5. Compare augmentation with and without the representation conversion at several small target-data sizes (10 min to 1 h), showing the benefit is largest when target data is tiny.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Before adding a treebank or silver corpus, measure how many of our target-domain lemmas, tags or constructions it fails to cover after conversion, and prioritize sources with the lowest residual OOV.
  • I2 (tree): When converting other UD dialects or treebanks, map unsupported relations or features to their closest Mova label instead of dropping the sentences, so imperfect sources still contribute.
  • I3 (tokens): A morphology or tagger model trained on related data could be constrained at inference by an English lexicon, so outputs always land on valid lemmas.
  • I4 (back): Generation from the meaning graph should target readable text directly, avoiding opaque intermediate forms that would need extra, error-prone conversion.
  • I5 (learn): Evaluate each silver-data or conversion trick on a ladder of gold-data sizes, with an "unconverted" control, to see where the trick actually pays off.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of California, BerkeleyBerkeley · United States175 works · 227,440 citations🎓Mohamed bin Zayed University of Artificial IntelligenceAbu Dhabi · United Arab Emirates124 works · 5,676 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSignal Processing · 452 works

🧭 Topics 3Research subjects and languages the work is about.Semitic languages · 2,908 worksSpeech & spoken language · 480 worksLow-resource languages · 267 works

Cite

Cite as Hellina Hailu Nigatu and Hanan Aldarmaki. 2025. Exploring Transliteration-Based Zero-Shot Transfer for Amharic ASR. Proceedings of the Sixth Workshop on African Natural Language Processing (AfricaNLP 2025).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.africanlp-1.10

@inproceedings{nigatu2025exploring,
  title = {{Exploring Transliteration-Based Zero-Shot Transfer for Amharic ASR}},
  author = {Hellina Hailu Nigatu and Hanan Aldarmaki},
  booktitle = {Proceedings of the Sixth Workshop on African Natural Language Processing (AfricaNLP 2025)},
  year = {2025},
  doi = {10.18653/v1/2025.africanlp-1.10},
  url = {https://doi.org/10.18653/v1/2025.africanlp-1.10}
}

References

24 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Siyuan Feng, Piotr ˙Zelasko, Laureano Moro-Vel´azquez, Ali Abavisani, Mark Hasegawa-Johnson, Odette Scharenborg, and Najim Dehak. 2021. find
  2. How Phonotactics Affect Multilingual and Zero-Shot ASR Performance. 2023. The Arabic Alphabet: Vowels. [Online; accessed 14. find
  3. Gabe Adugna. 2023. Research: Language Learning -Amharic: Home. [Online; accessed 14. find
  4. M. Gales, K. Knill, A. Ragni, and S. Rath. 2014. Speech recognition and keyword spotting for low-resource languages: Babel project research at CUED. Workshop on Spoken Language Technologies for Under-resourced Languages. find
  5. Hanan Aldarmaki and Ahmad Ghannam. 2023. Diacritic Recognition Performance in Arabic ASR. In Proc. INTERSPEECH 2023, pages 361–365.
  6. Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4218–4222, Marseille, France. European Language Resources Association. ↗
  7. Heting Gao, Junrui Ni, Yang Zhang, Kaizhi Qian, Shiyu Chang, and Mark Hasegawa-Johnson. 2021. Zero-Shot Cross-Lingual Phonetic Recognition with External Language Embedding. find
  8. José Ignacio Hualde. 2005. The sounds of spanish. find
  9. Jacob Kahn, Ann Lee, and Awni Hannun. Self-Training for End-to-End Speech Recognition. In ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 04–08. IEEE. find
  10. Israel Abebe Azime and Nebil Mohammed. 2021. An amharic news text classification dataset. In AfricaNLP. ↗
  11. Shreya Khare, Ashish R. Mittal, Anuj Diwan, Sunita Sarawagi, P. Jyothi, and Samarth Bharadwaj. 2021. Low Resource ASR: The Surprising Effectiveness of High Resource Transliteration. Interspeech. find
  12. Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021. VoxPopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 993–1003, Online. Association for Computational Linguistics. ↗
  13. Ajinkya Kulkarni, Atharva Kulkarni, Sara Abedal-mon’em Mohammad Shatnawi, and Hanan Aldarmaki. 2023. ClArTTS: An Open-Source Classical Arabic Text-to-Speech Corpus. In Proc. INTER-SPEECH 2023, pages 5511–5515.
  14. Wolf Leslau. 2000. Introductory grammar of Amharic, volume 21. Otto Harrassowitz Verlag. find
  15. Qiantong Xu, Alexei Baevski, and Michael Auli. 2022. Simple and Effective Zero-shot Cross-lingual Phoneme Recognition. find
  16. Xinjian Li, Florian Metze, David R. Mortensen, Alan W. Black, and Shinji Watanabe. 2022. ASR2K: Speech Recognition for Around 2000 Languages without Audio. find
  17. Piotr Żelasko, Laureano Moro-Velázquez, Mark Hasegawa-Johnson, Odette Scharenborg, and Najim Dehak. 2020. That Sounds Familiar: An Analysis of Phonetic Representations Transfer Across Languages. find
  18. Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2023. Scaling Speech Technology to 1,000+ Languages. arXiv. ↗
  19. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust Speech Recognition via Large-Scale Weak Supervision. arXiv. ↗
  20. Aku Rouhe, Anja Virkkunen, Juho Leinonen, and Mikko Kurimo. 2022. Low Resource Comparison of Attention-based and Hybrid ASR Exploiting wav2vec 2.0. In Proc. Interspeech 2022, pages 3543– 3547. find
  21. Sebastian Spiegler, Andrew van der Spuy, and Peter A. Flach. 2010. Ukwabelana - An open-source morphological Zulu corpus. ACL Anthology, pages 1020– 1028. find
  22. Martha Yifiru Tachbelie, Solomon Teferra Abate, and Laurent Besacier. 2014. Using different acoustic, lexical and language modeling units for ASR of an under-resourced language – Amharic. Speech Communication, 56:181–194. ↗
  23. Martha Yifiru Tachbelie, Solomon Teferra Abate, and Tanja Schultz. 2020a. Analysis of GlobalPhone and Ethiopian languages speech corpora for multilingual ASR. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4152–4156, Marseille, France. European Language Resources Association. find
  24. Martha Yifiru Tachbelie, Solomon Teferra Abate, and Tanja Schultz. 2020b. DNN-based multilingual automatic speech recognition for Wolaytta using Oromo speech. In Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced languages (SLTU) and Collaboration and Computing for Under-Resourced Languages (CCURL), pages 265–270, Marseille, France. European Language Resources association. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nigatu-2025-exploring-transliteration-based-zero-shot/