Elamur Research⌕
☰
home / papers / 2026 / OMAN-SPEECH: A Multi-Layer Annotated Speech Corpus for Omani Arabic Dialects

OMAN-SPEECH: A Multi-Layer Annotated Speech Corpus for Omani Arabic Dialects

Paper · 2026 · Rayyan S. Al Khadhuri, Firas Al Mahrouqi, Salim Al Mandhari et al.

Summary

Omani Arabic has very little speech data. It is usually folded into broad Gulf Arabic collections, or represented only by Muscat, which the paper argues is a linguistically mixed capital and a poor stand-in for the country. The authors build OMAN-SPEECH, a corpus of about 40.5 hours of natural and partly guided talk recorded by 32 speakers across 11 Wilayats. Each Wilayat is labelled by lifestyle (Hadari, Badu or Mixed) following the typology of Alkathiri (2024).

Speakers recorded at home in small groups of people from the same community, plus individual monologues, using topics chosen from a list. An in-house tool supported the annotation in stages:

  1. ElevenLabs Scribe v1 produced draft transcripts.
  2. Native annotators corrected them with speaker diarization.
  3. Annotators then added sentence-level English translation and IPA transcription.

The paper reports no ASR results. Its contribution is the resource itself and the collection design. The authors argue it can serve as a benchmark for dialect-aware ASR, translation and dialect identification on an under-covered Arabic variety.

Key points

  • Size and speakers: about 40.5 hours in total, with roughly 22 hours of group conversation and about 18–18.5 hours of monologue. There are 32 speakers (14 male, 18 female) in 11 groups.
  • Regional coverage: the 11 Wilayats include Musandam, Shinas, Nizwa (the largest at 9 h), Salalah and Al Jazer (the only Badu-labelled Wilayat, 1.5 h). Muscat is not among them.
  • Recording protocol: each group recorded a three-hour conversation, and each speaker a one-hour monologue. Topics covered areas such as daily customs, food, folklore and social or political debate.
  • Annotation layers: Arabic transcription, English translation and IPA, all aligned to corrected, time-stamped segments. IPA is meant to capture features that Arabic spelling hides, such as /q/ pronounced as [g].
  • Planned work: annotation of the full corpus is not yet complete. The authors plan baselines with Whisper and Wav2Vec 2.0, a public release, and coverage of the remaining Wilayats.

Interesting ideas

  • I1. Sample data along a social axis (settled versus nomadic lifestyle), not just by location, and refuse to let the capital stand in for the whole language; this exposes variation a plain regional split would hide.
  • I2. Record speakers only alongside peers from their own community, so they drift less toward the formal standard while being observed; the design of elicitation shapes how natural the data is.
  • I3. Let a machine produce a rough draft, have humans correct it, and only then add further layers tied to the corrected segments; later work never rests on unverified boundaries.
  • I4. Keep several parallel views of each sentence (native script, translation, phonetic form), because each one records information the others lose.
  • I5. Use an explicit "Mixed" label instead of forcing every region into a binary class, which keeps the labels honest where categories blur.

For Mova

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

  • I1 (learn): Build evaluation sets stratified by register (children's tales, school texts, informal dialogue) and report scores per stratum, so a gain on one genre cannot hide a loss on another.
  • I2 (learn): When an LLM teacher generates practice text, prompt it inside a consistent in-genre scenario rather than in neutral style, so the silver data keeps genuine informal constructions.
  • I3 (learn): Freeze the corrected UD tree before attaching frames and roles, keyed to stable token IDs, so repairs to the tree do not silently break the semantic layer above it.
  • I4 (back): Store source sentence, meaning graph and regenerated text side by side per sentence, so that round-trip losses in the compressor can be located exactly.
  • I5 (world): Allow ontology and entity-state labels to take an explicit mixed or undetermined value, rather than forcing a choice the text does not support.

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

Also on the byline: University of Technology and Applied Sciences.

 Fields, topics and concepts

📚 Shelves 4The 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 works

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

Cite

Cite as Rayyan S. Al Khadhuri, Firas Al Mahrouqi, Salim Al Mandhari, Amir Azad Al-Kathiri, Omar Said Alshahri, Ghassab Mansoor Alsaqr, Badri Abdulhakim Mudhsh and Tarek Fatnassi. 2026. OMAN-SPEECH: A Multi-Layer Annotated Speech Corpus for Omani Arabic Dialects. Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.abjadnlp-1.31

@inproceedings{khadhuri2026oman,
  title = {{OMAN-SPEECH: A Multi-Layer Annotated Speech Corpus for Omani Arabic Dialects}},
  author = {Rayyan S. {Al Khadhuri}, Firas {Al Mahrouqi}, Salim {Al Mandhari}, Amir Azad Al-Kathiri, Omar Said Alshahri, Ghassab Mansoor Alsaqr, Badri Abdulhakim Mudhsh and Tarek Fatnassi},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.31},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.31}
}

References

17 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Muhammad Abdul-Mageed, Hassan Alhuzali, and Mohamed Elaraby. 2018. You tweet what you speak: A city-level dataset of Arabic dialects. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).
  2. Kathrein Abu Kwaik, Motaz Saad, Stergios Chatzikyriakidis, and Simon Dobnik. 2018. Shami: A corpus of Levantine Arabic dialects. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).
  3. Rashid Al-Balushi. 2017. Omani arabic: More than a dialect. Macrolinguistics, 4:80–125. ↗
  4. Khoula Al-Kharusi and Abdurahman AAlAbdulsalam. 2023. Machine translation of Omani Arabic dialect from social media. In Proceedings of ArabicNLP 2023, pages 302–309, Singapore (Hybrid). Association for Computational Linguistics. Southern Denmark University, Odense, Denmark. Association for Computational Linguistics.
  5. Fatma Zahra Besdouri, Inès Zribi, and Lamia Hadrich Belguith. 2024. Arabic automatic speech recognition: Challenges and progress. Speech Communication, 163:103110. ↗
  6. Houda Bouamor, Nizar Habash, Mohammad Salameh, Wajdi Zaghouani, Owen Rambow, Dana Abdulrahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, and Kemal Oflazer. 2018. The MADAR Arabic dialect corpus and lexicon. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). find
  7. Soumia Bougrine, Aicha Chorana, Abdallah Lakhdari, and Hadda Cherroun. 2017. Toward a web-based speech corpus for Algerian dialectal Arabic varieties. In Proceedings of the Third Arabic Natural Language Processing Workshop, pages 138–146, Valencia, Spain. Association for Computational Linguistics. ↗
  8. Amira Dhouib, Achraf Othman, Oussama El Ghoul, Mohamed Koutheair Khribi, and Aisha Al Sinani. 2022. Arabic automatic speech recognition: A systematic literature review. Applied Sciences, 12(17). ↗
  9. Amirbek Djanibekov, Hawau Olamide Toyin, Raghad Alshalan, Abdullah Alatir, and Hanan Aldarmaki. 2025. Dialectal coverage and generalization in Arabic speech recognition. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29490–29502, Vienna, Austria. Association for Computational Linguistics. ↗
  10. Ashraf Elnagar, Sane M. Yagi, Ali Bou Nassif, Ismail Shahin, and Said A. Salloum. 2021. Systematic literature review of dialectal arabic: Identification and detection. IEEE Access, 9:31010–31042. ↗
  11. Charles A. Ferguson. 1959. 15(2):325–340. Diglossia. find
  12. WORD, Nizar Y. Habash. 2010. Introduction to Arabic natural language processing, 1 edition, volume 3 of Synthesis Lectures on Human Language Technologies. Morgan and Claypool Publishers. find
  13. Ahmed Ali, Stephan Vogel, and Steve Renals. 2017. Speech recognition challenge in the wild: Arabic mgb-3. Preprint, arXiv:1709.07276. ↗
  14. Clive Holes. 1989. Towards a dialect geography of oman. Bulletin of the School of Oriental and African Studies, 52(3):446–462. ↗
  15. Amir Alkathiri. 2024. Toward a new geographic classification of dialectal groups in the sultanate of oman. Al-Adab Journal, 149:71–98. find
  16. Abdul Waheed, Bashar Talafha, Peter Sullivan, Abdel-Rahim Elmadany, and Muhammad Abdul-Mageed. 2023. VoxArabica: A robust dialect-aware Arabic speech recognition system. In Proceedings of ArabicNLP 2023, pages 441–449, Singapore (Hybrid). Association for Computational Linguistics.
  17. Noureddine Atouf, Elsayed Issa, and Said Ouzbayr. 2025. Tachelhiyt-Darija: a parallel speech corpus for two underrepresented languages. In Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP-2025), pages 379–384,

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/khadhuri-2026-oman-speech-multi-layer-annotated/