Elamur Research⌕
☰
home / papers / 2026 / Fleurs-Badini: Translation and Recording Fleurs Dataset for Badini Variant of Northern Kurdish

Fleurs-Badini: Translation and Recording Fleurs Dataset for Badini Variant of Northern Kurdish

Summary

Badini, a Northern Kurdish variety spoken mainly in Iraq's Duhok governorate and usually written in a modified Arabic script, has been missing from the FLEURS benchmark. Kurdish speech resources have concentrated on Central Kurdish, and only recently on Kurmanji through FLEURS-Kobani. To fill this gap, the authors built FLEURS-Badini. Students at the University of Duhok translated 2,000 English FLORES sentences directly into Badini, and faculty reviewed the translations. Native speakers then read the sentences aloud through a web platform on their own devices, and every recording was checked by hand. The final corpus contains about 15.7 hours of speech paired with English text, split so that no speaker appears in more than one partition. Baseline experiments show the task is hard. The best ASR system, a W2V-BERT CTC model, still makes roughly half its word-level predictions wrong, and a fine-tuned Whisper model reaches only single-digit BLEU for Badini-to-English speech translation. The resource matters because it is the first benchmark for speech recognition and speech translation in this dialect.

Key points

  • Dataset size and splits: FLEURS-Badini has 5,224 utterances (15h40m) from 45 speakers.
    • Train: 2,022 utterances, 5h47m, 16 speakers
    • Dev: 1,165 utterances, 3h36m, 14 speakers
    • Test: 2,037 utterances, 6h17m, 15 speakers
    • Validation removed 1,078 recordings.
  • Translation process: 50 native-speaker students handled the translation in two stages. They dealt with gaps in specialized vocabulary by borrowing, transliterating or explaining terms, and they adjusted English SVO order to Kurdish SOV order.
  • ASR results: W2V-BERT CTC performed best, with CER of 13.69 (dev) and 14.11 (test) and about 55% WER on test. Its training data was Common Voice Kurmanji transliterated into Arabic script plus synthetic data.
    • Omnilingual LLM 1B reached about 60% WER, and the 3B model did worse, partly because it misidentified the language or script.
    • Whisper fine-tuned on FLEURS-Badini reached only about 39% CER.
  • S2TT results: Whisper Large V3 fine-tuned on the train split scored BLEU 6.13 and chrF++ 29.39 on dev, and BLEU 5.24 and chrF++ 29.57 on test.
  • Availability: The dataset is on Hugging Face (BadiniSpeechNLP/FLEURS-badini) under the CC-BY-NC-ND-4.0 license.

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

Fields: Linguistics & languages · Machine translation

Topics: Machine translation

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.iwslt-1.14

@inproceedings{mohammadamini2026fleurs,
  title = {{Fleurs-Badini: Translation and Recording Fleurs Dataset for Badini Variant of Northern Kurdish}},
  author = {Mohammadamini, Mohammad and Tayib, Dilgash Mohammed Salih and Abdulazeez, Dezheen H. and Mohammed, Barzan Hussein and Sadeeq, Imad Saeed and Mohammed, Aveen Jalal and Melhum, Amera Ismail and Dheyab, Abuobaida Abdullah},
  booktitle = {Proceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.iwslt-1.14},
  url = {https://doi.org/10.18653/v1/2026.iwslt-1.14}
}

References

19 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. David Ifeoluwa Adelani, Victor Agostinelli, Antonios Anastasopoulos, Luisa Bentivogli, Ondřej Bojar, Sebastien Bratières, Marine Carpuat, Roldano Cattoni, Mauro Cettolo, Lizhong Chen, Marcello Federico, Marco Gaido, Mahendra Gupta, HyoJung Han, Ali Hatami, David Javorský, Yejin Jeon, Marek Kaszteln… find
  2. Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M. Tyers, and Gregor Weber. 2020. Common voice: A massively multilingual speech corpus. In Proceedings of the 12th Language Resources and Evaluation Conference (LREC), pages 4218–4222. find our card
  3. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. find our card
  4. Model Omnilingual LLM 1B Omnilingual LLM 3B W2V-BERT CTC Whisper Large V3 Baseline Whisper CER (Dev) 14.81 20.97 13.69 15.71 39.33 WER (Dev) 60.06 63.97 54.46 60.13 70.87 CER (Test) 14.75 22.34 14.11 16.50 39.05 WER (Test) 60.20 64.47 55.07 62.02 72.16 *Table 2: ASR results across different models*… find
  5. Alexis Conneau, Ming Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Marc’Aurelio Ranzato, and Veselin Stoyanov. 2022. Fleurs: Few-shot learning evaluation of universal representations of speech. In IEEE Spoken Language Technology Workshop (SLT). find our card
  6. Eva Eppler and Birgit Benedikt. 2017. Kurdish linguistics. In The Routledge Handbook of Kurdish Linguistics. Routledge. find
  7. Naman Goyal, Jianfeng Gao, Vishrav Chaudhary, Peng-Jen Chen, Guillaume Wenzek, Da Ju, Saurabh Krishnan, Marc’Aurelio Ranzato, and Angela Fan. 2021. The flores-101 evaluation benchmark for low-resource and multilingual machine translation. In ACL / TACL. find
  8. Geoffrey Haig and Ergin Öpengin. 2014. Kurdish: A critical research overview. Iranian Studies, 47(2):191–225. find
  9. Daban Q. Jaff and Mohammad Mohammadamini. 2026. Fleurs-kobani: Extending the fleurs dataset for northern kurdish. Preprint, arXiv:2603.29892. — [arXiv:2603.29892](https://arxiv.org/abs/2603.29892) find
  10. Mohammad Mohammadamini, Daban Jaff, Sara Jamal, Ibrahim Ahmed, Hawkar Omar, Darya Sabr, Marie Tahon, and Antoine Laurent. 2025. Kuvost: A large-scale human-annotated English to Central Kurdish speech translation dataset driven from English common voice. In Proceedings of the 22nd International Confe… find
  11. Mohammad Mohammadamini, Aveen Jalal Mohammed, Barzan Hussein Mohammed, Dezheen H Abdulazeez, Imad Saeed Sadeeq, Dilgash Mohammed, Salih Tayib, Amera Ismail Melhum, and Abuobaida Abdullah Dheyab. 2026. Exploring the Reusability of Northern Kurdish Resources for Badini Speech OmnilingualASR, Gil Keren… — [arXiv:2511.09690](https://arxiv.org/abs/2511.09690) find
  12. Martin Popel and Ondřej Bojar. 2020. Transforming machine translation: The flores benchmark. In ACL. find
  13. Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Massively multilingual asr: 50 languages, 1 model. In Interspeech. find
  14. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In ICML. find our card
  15. Jaffer Sheyholislami. 2015. Language varieties of the kurds. In Wolfgang Taucher, Mathias Vogl, and Peter Webinger, editors, The Kurds: History, Religion, Language, Politics, pages 30–51. Austrian Federal Ministry of the Interior, Vienna, Austria. find
  16. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NeurIPS. find
  17. Hadi Veisi, Hawre Hosseini, Mohammad MohammadAmini, Wirya Fathy, and Aso Mahmudi. 2022. Jira: A central kurdish speech recognition system, designing and building speech corpus and pronunciation lexicon. Language Resources and Evaluation, 56:917–941. find
  18. Thomas Zenkel, Ramon Sanabria, Florian Metze, and Alex Waibel. 2020. Blstm-hmm hybrid systems for low-resource speech recognition. In Interspeech. find
  19. Ergin Öpengin. 2021. The history of kurdish and the development of literary kurmanji. In Hamit Bozarslan, Cengiz Gunes, and Veli Yadirgi, editors, The Cambridge History of the Kurds, pages 603–632. Cambridge University Press. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/mohammadamini-2026-fleurs-badini-translation-recording-fleurs/