Elamur Research⌕
☰
home / papers / 2026 / ADAPT–MTU HAI at IWSLT2026: Robust Cascaded Speech Translation for Bhojpuri–Hindi and Irish–English

ADAPT–MTU HAI at IWSLT2026: Robust Cascaded Speech Translation for Bhojpuri–Hindi and Irish–English

Summary

This paper presents WhiNN-ST, the submission from ADAPT–MTU HAI to the IWSLT 2026 shared task on speech translation in low-resource settings. It covers two pairs, Bhojpuri→Hindi and Irish→English, where scarce data and weak ASR support make end-to-end models unreliable. The authors built a modular cascade: Whisper-based (or language-specific) ASR feeds the NLLB-200 distilled 600M translation model. They added librosa audio preprocessing (16 kHz mono), heuristic script normalization, and a custom manifest pipeline that repairs inconsistent TSV metadata. They compared ASR front-ends and two routing options: direct translation and pivoting through a high-resource language. With a weaker Whisper-medium front-end, pivoting through English helped. Once Whisper-large-v3 and a genuine NLLB direct route were used, direct translation won on all surface metrics. This suggests pivoting mainly compensates for a weak direct path. For Irish, an Irish-specific Wav2Vec2 model eliminated empty transcriptions. The study shows that careful ASR choice and light engineering, without extra training, can yield robust low-resource speech translation.

Key points

  • Best Bhojpuri→Hindi setup was Whisper-large-v3 with direct NLLB. On the full development set (1056 segments), it reached BLEU 25.59, chrF++ 42.48, and TER 63.83, beating pivot and copy baselines.
  • The Vakyansh Bhojpuri ASR failed badly on a 50-sample subset, scoring BLEU 0.33 with direct translation. Whisper-large-v3 scored 22.82 on the same subset.
  • For Irish→English, swapping Whisper-large-v3 for Wav2Vec2 Irish raised non-empty ASR coverage from 94.8% to 100% on 722 test segments. Repetition rates stayed low, and SeamlessM4T-v2 could not be run successfully.
  • Test sets contain 252 segments for Bhojpuri (≈23h of data) and 722 segments for Irish (≈206h). There was no Irish dev set with references, so evaluation relied on coverage, repetition, and length-ratio diagnostics, and WER could not be computed for either pair.
  • Final submissions used direct NLLB as primary for both pairs. The contrastive runs pivoted via English for Bhojpuri and via Hindi for Irish.

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-1.6

@inproceedings{sonawane2026adapt,
  title = {{ADAPT–MTU HAI at IWSLT2026: Robust Cascaded Speech Translation for Bhojpuri–Hindi and Irish–English}},
  author = {Sonawane, Pournima and Afli, Haithem},
  booktitle = {Proceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.iwslt-1.6},
  url = {https://doi.org/10.18653/v1/2026.iwslt-1.6}
}

References

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

  1. Yao-Fei Cheng, Hung-Shin Lee, and Hsin-min Wang. 2021. Allost: Low-resource speech translation without source transcription. In Proceedings of Interspeech 2021, pages 2252–2256. find
  2. 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
  3. Akshat Govil, C Ravindra Reddy, Haren A, and Manju Venugopalan. 2026. Fine-tuning meta’s nllb-200 model for telugu–english neural machine translation and real-time wikihow deployment via android app. In 2026 International Conference on Computing, Communication, Control and Cyber-Physical Systems (I5… find
  4. Séamus Lankford. 2024. Enhancing neural machine translation of low-resource languages: Corpus development, human evaluation and explainable ai architectures. Preprint, arXiv:2403.01580. find
  5. Haithem Afli, Loïc Barrault, and Holger Schwenk. 2013. Multimodal comparable corpora as resources for extracting parallel data: Parallel phrases extraction. In Proceedings of the Sixth International Joint Conference on Natural Language Processing. find
  6. Séamus Lankford, Haithem Afli, and Andy Way. 2022. Human evaluation of english–irish transformer-based neural machine translation. Information, 13(7):309. find
  7. Marko Avila and Josep Crego. 2025. SYSTRAN @ IWSLT 2025 low-resource track. In Proceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), pages 324–332, Vienna, Austria (in-person and online). Association for Computational Linguistics. find
  8. Chutong Meng and Antonios Anastasopoulos. 2025. GMU systems for the IWSLT 2025 low-resource speech translation shared task. In Proceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), pages 289–300, Vienna, Austria (in-person and online). Association for Computat… find
  9. Sameer Bansal, Herman Kamper, Karen Livescu, Adam Lopez, and Sharon Goldwater. 2018. Low-resource speech-to-text translation. In Proceedings of Interspeech 2018, pages 1298–1302. find
  10. Yasmin Moslem. 2024. Leveraging synthetic audio data for end-to-end low-resource speech translation. In Proceedings of the 21st International Conference on Spoken Language Translation (IWSLT 2024), pages 265–273, Bangkok, Thailand (in-person and online). Association for Computational Linguistics. find
  11. Kaushal Bhogale, Abhigyan Raman, Tahir Javed, Sumanth Doddapaneni, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2023. Effectiveness of mining audio and text pairs from public data for improving asr systems for low-resource languages. In ICASSP 2023 - 2023 IEEE International Conference o… find
  12. NVIDIA. 2024. Whisper large v3 - nvidia ai foundation models. https://build.nvidia.com/ openai/whisper-large-v3. Accessed: 2026. ↗
  13. OpenAI. 2023. Whisper large v3 model. https://huggingface.co/openai/whisper-large-v3. Accessed: 2026. ↗
  14. Juan N. Pava, Caroline Meinhardt, Haifa Badi Uz Zaman, Toni Friedman, Sang T. Truong, Daniel Zhang, Elena Cryst, Vukosi Marivate, and Sanmi Koyejo. 2025. Mind the language gap: Mapping the challenges of llm development in low-resource language contexts. find
  15. Nathaniel Romney Robinson, Kaiser Sun, Cihan Xiao, Niyati Bafna, Weiting Tan, Haoran Xu, Henry Li Xinyuan, Ankur Kejriwal, Sanjeev Khudanpur, Kenton Murray, and Paul McNamee. 2024. find
  16. JHU IWSLT 2024 dialectal and low-resource system description. In Proceedings of the 21st International Conference on Spoken Language Translation (IWSLT 2024), pages 140–153, Bangkok, Thailand (in-person and online). Association for Computational Linguistics. find
  17. Shashwat Upadhyay. 2026a. find
  18. Iwslt 2026 bhojpuri-hindi dataset repository. https://github.com/ shashwatup9k/iwslt2026_bho-hi. ↗
  19. Shashwat Upadhyay. 2026b. find
  20. Iwslt 2026 irish-english dataset repository. https://github.com/ shashwatup9k/iwslt2026_ga-eng. ↗
  21. Vistatec. 2025. How to overcome the need for data for low-resource languages. find
  22. Arlisa Yuliawati, Ika Alfina, and Indra Budi. 2025. From corpus to benchmark: Evaluating pretrained language models for indonesian-javanese krama translation. In 2025 International Conference on Asian Language Processing (IALP), pages 199–204. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/sonawane-2026-adapt-mtu-hai-iwslt2026-robust/