Elamur Research⌕
☰
home / papers / 2026 / LIUM Submission for IWSLT 2026 Low-resource Speech Translation Track

LIUM Submission for IWSLT 2026 Low-resource Speech Translation Track

Summary

The paper addresses the scarcity of human-annotated speech translation data for Central Kurdish–English (ckb–en) in the IWSLT 2026 low-resource track. Its test set, Kurdish-Commute, contains spontaneous media speech. The authors compare two ways of building end-to-end training triplets without manual labeling.

  1. Pseudo-labeling (ASR+MT): about 4,300 hours of Kurdish audiobooks are segmented with VAD, transcribed by a fine-tuned Seamless Large V2, and translated by a fine-tuned NLLB 1.3B. The output is then filtered by speaking rate, duration, ASR confidence, repetition-based hallucination checks, length ratio and named-entity share.
  2. Speech synthesis (TTS+MT): F5-TTS voices voice either existing bitext or raw news text, which is then machine-translated.

A transformer Fairseq-S2T model trained from scratch on the pseudo-labeled data performed best. Its BLEU roughly matches a Seamless+NLLB cascade, while an end-to-end system avoids the cascade's added latency. Synthetic speech generalized poorly to spontaneous speech, likely because it lacks disfluencies, prosodic variation, code-switching and dialect diversity. Per the IWSLT 2026 Findings, these systems ranked first for ckb–en in both the constrained and open settings.

Key points

  • Pseudo-labeled corpus: filtering kept 1.71M triplets, about 3,231 h of speech and ~22.66M English tokens. The ASR was fine-tuned on 177 h from Common Voice 18 and Asosoft, and the MT on 222k ckb–en sentence pairs.
  • Best system (Fairseq-S2T, 268M parameters, pseudo-labeled data):
    • 25.73 BLEU on dev and 21.09 BLEU on test.
    • ChrF++ of 54.28 on dev and 49.48 on test.
    • For comparison, Whisper Large V3 trained on the same data scored about 18 BLEU.
  • Synthetic data results:
    • Voicing existing bitext (mostly Kuvost, 514 h) gave about 10.4 test BLEU.
    • The raw-text TTS+MT route gave 12.69 test BLEU.
    • Combining synthetic and pseudo-labeled data gave 16.24 test BLEU, below pseudo-labeled data alone.
  • Reference points:
    • The constrained baseline (Whisper Large V3 fine-tuned only on the COMMUTE-Kurdish training split) reached 18.85 BLEU on dev and 14.94 on test.
    • The Seamless+NLLB cascade reached 24.77 BLEU on dev and 22.39 on test.
  • ASR comparison:
    • Fine-tuned Seamless-L-V2 was clearly the strongest model, at 17.31 WER on dev and 19.76 on test.
    • Fairseq-S2T trained on its pseudo-labels came close, at 24.47 WER on dev and 25.78 on test.
    • The non-fine-tuned Omnilingual baselines reached roughly 37–47 WER.

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-1.15

@inproceedings{mohammadamini2026lium,
  title = {{LIUM Submission for IWSLT 2026 Low-resource Speech Translation Track}},
  author = {Mohammadamini, Mohammad and Tahon, Marie},
  booktitle = {Proceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.iwslt-1.15},
  url = {https://doi.org/10.18653/v1/2026.iwslt-1.15}
}

References

26 references, 3 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. Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Paul-Ambroise Duquenne, Hady Elsahar, Hongyu Gong, Kevin Heffernan, et. all., and SEAMLESS Communication Team. 2025. Joint speech and text machine translation for up to 100 languages. Nature, 637(8046):587–593. find
  3. Martijn Bartelds, Nay San, Bradley McDonnell, Dan Jurafsky, and Martijn Wieling. 2023. Making more of little data: Improving low-resource automatic speech recognition using data augmentation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long P… find
  4. Kaushal Santosh Bhogale, Deovrat Mehendale, Niharika Parasa, Sathish Kumar Reddy G, Tahir Javed, Pratyush Kumar, and Mitesh M. Khapra. 2024. Empowering low-resource language asr via large-scale pseudo labeling. In Interspeech 2024, pages 2519– 2523. find
  5. Yushen Chen, Zhikang Niu, Ziyang Ma, Keqi Deng, Chunhui Wang, Jian Zhao, Kai Yu, and Xie Chen. 2025. F5-TTS: A fairytaler that fakes fluent and faithful speech with flow matching. Preprint, arXiv:2410.06885. find
  6. Yu-An Chung, Yu Zhang, Wei Han, Chung-Cheng Chiu, James Qin, Ruoming Pang, and Yonghui Wu. 2021. W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training. Preprint, arXiv:2108.06209. — [arXiv:2108.06209](https://arxiv.org/abs/2108.06209) find
  7. Alexis Conneau, Min Ma, Simran Khanuja, Yu Zhang, Vera Axelrod, Siddharth Dalmia, Jason Riesa, Clara Rivera, and Ankur Bapna. 2023. Fleurs: Few-shot learning evaluation of universal representations of speech. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 798–805. find our card
  8. Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman, and 2… find
  9. Yaroslav Getman, Tamas Grosz, Katri Hiovain-Asikainen, and Mikko Kurimo. 2024. Exploring adaptation techniques of large speech foundation models for low-resource asr: a case study on northern sámi. In Interspeech 2024, pages 2539–2543. find
  10. Yosuke Higuchi, Niko Moritz, Jonathan Le Roux, and Takaaki Hori. 2022. Momentum pseudo-labeling: Semi-supervised asr with continuously improving pseudo-labels. IEEE Journal of Selected Topics in Signal Processing, 16(6):1424–1438. find
  11. Dongseong Hwang, Khe Chai Sim, Zhouyuan Huo, and Trevor Strohman. 2022. Pseudo label is better than human label. In Interspeech 2022, pages 1421–1425. find
  12. Sameer Khurana, Niko Moritz, Takaaki Hori, and Jonathan Le Roux. 2021. Unsupervised domain adaptation for speech recognition via uncertainty driven self-training. pages 6553–6557. find
  13. Zhaolin Li, Yining Liu, Danni Liu, Tuan Nam Nguyen, Enes Yavuz Ugan, Tu Anh Dinh, Carlos Mullov, Alexander Waibel, and Jan Niehues. 2025. KIT’s low-resource speech translation systems for IWSLT2025: System enhancement with synthetic data and model regularization. In Proceedings of the 22nd Internati… find
  14. Christoph Minixhofer, Ondřej Klejch, and Peter Bell. 2025. Scaling Laws for Synthetic Speech for Model Training. In Interspeech 2025, pages 3189–3193. find
  15. Tomoya Mizumoto, Atsushi Kojima, Yusuke Fujita, Lianbo Liu, and Yui Sudo. 2025. Is Synthetic Data Truly Effective for Training Speech Language Models? In Interspeech 2025, pages 1808–1812. find
  16. Mohammad Mohammadamini, Daban Jaff, Sara Jamal, Ibrahim Ahmed, Hawkar Omar, Darya Sabr, Marie Tahon, and Antoine Laurent. 2025a. Kuvost: A large-scale human-annotated English to Central Kurdish speech translation dataset driven from English common voice. In Proceedings of the 22nd International Conf… find
  17. Mohammad Mohammadamini, Meysam Shamsi, and Marie Tahon. 2026. Central Kurdish text-to-speech and its application in speech-to-text translation. In Language Resources and Evaluation Conference (LREC), Palma, Spain. find our card
  18. Mohammad Mohammadamini, Aghilas Sini, Marie Tahon, and Antoine Laurent. 2025b. Scaling pseudo-labeling data for end-to-end low-resource speech translation (the case of Kurdish language). In Interspeech 2025, pages 898–902. find
  19. Rabindra Nath Nandi, Mehadi Menon, Tareq Muntasir, Sagor Sarker, Quazi Sarwar Muhtaseem, Md. Tariqul Islam, Shammur Chowdhury, and Firoj Alam. 2023. Pseudo-labeling for domain-agnostic Bangla automatic speech recognition. pages 152–162. find
  20. OmnilingualASR, Gil Keren, Artyom Kozhevnikov, Yen Meng, Christophe Ropers, Matthew Setzler, Skyler Wang, Ife Adebara, Michael Auli, Can Balioglu, Kevin Chan, Chierh Cheng, Joe Chuang, Caley Droof, Mark Duppenthaler, Paul-Ambroise Duquenne, Alexander Erben, Cynthia Gao, Gabriel Mejia Gonzalez, and 1… — [arXiv:2511.09690](https://arxiv.org/abs/2511.09690) find
  21. Yu Pu, Xiaoqian Liu, Guangyu Zhang, Zheng Yan, Wei-Qiang Zhang, and Xie Chen. 2025. Empowering Large Language Models for End-to-End Speech Translation Leveraging Synthetic Data. In Interspeech 2025, pages 26–30. find
  22. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356. find our card
  23. Minh Tran, Debjyoti Paul, Yutong Pang, Laxmi Pandey, Jinxi Guo, Ke Li, Shun Zhang, Xuedong Zhang, and Xin Lei. 2025. R2S: Real-to-Synthetic Representation Learning for Training Speech Recognition Models on Synthetic Data. In Interspeech 2025, pages 3194–3198. find
  24. 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(3):917–941. find
  25. Hadi Veisi, Mohammad MohammadAmini, and Hawre Hosseini. 2020. Toward kurdish language processing: Experiments in collecting and processing the asosoft text corpus. Digital Scholarship in the Humanities, 35(1):176–193. find
  26. Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, and Juan Pino. 2020. Fairseq S2T: Fast speech-to-text modeling with fairseq. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Na… find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/mohammadamini-2026-lium-submission-iwslt-2026-low/