Elamur Research⌕
☰
home / papers / 2026 / The CUHKSZ System for the IWSLT 2026 Low-Resource Speech-to-Text Task

The CUHKSZ System for the IWSLT 2026 Low-Resource Speech-to-Text Task

Summary

Fine-tuning a single multilingual speech translation model on several low-resource languages at once often causes negative transfer: gradients from different languages push shared weights in opposite directions. For the IWSLT 2026 low-resource speech-to-text track, the CUHKSZ team (submitting as SLC) addresses this with Gradient-Driven Parameter Sharing (GDPS). Starting from SeamlessM4T-Medium, they measure per-language gradient statistics using three methods: pairwise cosine similarity with Ward clustering, a self-versus-cross conflict score, and joint SVD/CCA. These statistics determine how languages are grouped, how much capacity remains shared, and which layer to split. The analysis points to the FFN2 block of encoder Layer 11, which is rebuilt as one shared path plus group-specific private paths. Training adds curriculum distillation from SeamlessM4T-v2-Large, with pseudo-labels gradually mixed with reference data. At inference, outputs from multiple checkpoints are reranked using prior BLEU and self-consistency. Across eight languages, GDPS helps most where gradient conflict is strong (Bemba, Hausa), while the outlier Central Kurdish and Yoruba gain mainly from reranking. The work offers a data-driven alternative to hand-designed parameter sharing.

Key points

  • Language setup: The system covers 8 directions: six official IWSLT 2026 directions (bem, ckb, gle, hau, ibo, yor) plus aeb (IWSLT 2022) and est (LoResMT). All submissions use the unconstrained condition, with roughly 9k–20k lines of data per language.
  • Gradient analysis findings:
    • Layer 11 FFN2 showed the highest cross-language subspace disagreement (0.174).
    • ckb is a clear outlier, with self-similarity of 0.509 versus about 0.21 for cross-language pairs.
    • The derived shared ratio was ρ=0.5, with K=3 groups for Medium and K=4 for LargeV2.
  • GDPS gains over unified fine-tuning (UFT-Med): bem +2.07 BLEU, hau +1.50, ibo +0.38. GDPS-Med also had a slightly higher average COMET (0.573 vs. 0.568).
  • Distillation and reranking:
    • The best curriculum stage mixes 50% pseudo-labeled and 50% reference data.
    • Prior-BLEU+Consensus reranking reached 14.85 average BLEU.
    • It lifted ckb from 8.38 to 20.26 and yor from 8.03 to 13.23.
  • Official blind test: yor reached 11.7 spBLEU. The ckb result was withdrawn, which the authors attribute to a submission formatting error. Within-cluster gains were hard to predict, ranging from +1.50 to −0.40 BLEU in the Niger-Congo group.

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-1.33

@inproceedings{sun2026cuhksz,
  title = {{The CUHKSZ System for the IWSLT 2026 Low-Resource Speech-to-Text Task}},
  author = {SUN, ruiyan and Li, Qingming and Nakamura, Satoshi},
  booktitle = {Proceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.iwslt-1.33},
  url = {https://doi.org/10.18653/v1/2026.iwslt-1.33}
}

References

23 references, 8 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. Milind Agarwal, Sweta Agrawal, Antonios Anastasopoulos, Luisa Bentivogli, Ondřej Bojar, Claudia Borg, Marine Carpuat, Roldano Cattoni, Mauro Cettolo, Mingda Chen, and 1 others. 2023. Findings of the IWSLT 2023 evaluation campaign. In Proceedings of the 20th International Conference on Spoken Langua… find our card
  3. Antonios Anastasopoulos, Loïc Barrault, Luisa Bentivogli, Marcely Zanon Boito, Ondřej Bojar, Roldano Cattoni, Anna Currey, Georgiana Dinu, Kevin Duh, Maha Elbayad, and 1 others. 2022. Findings of the IWSLT 2022 evaluation campaign. In Proceedings of the 19th International Conference on Spoken Langu… find our card
  4. Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, and 1 others. 2023a. Seamless: Multilingual expressive and streaming speech translation. arXiv preprint arXiv:2312.05187. find our card
  5. Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, and 1 others. 2023b. Seamlessm4t: Massively multilingual & multimodal machine translation. arXiv preprint arXiv:2308.11596. find our card
  6. Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, and Dumitru Erhan. 2016. Domain separation networks. In Advances in neural information processing systems, pages 343–351. find
  7. Xin Dong, Ruize Wu, Chao Xiong, Hai Li, Lei Cheng, Yong He, and 1 others. 2022. Gdod: Effective gradient descent using orthogonal decomposition for multi-task learning. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 386–395. find
  8. Isaac Gerber. 2025. Attention is not all you need: The importance of feedforward networks in transformer models. arXiv preprint arXiv:2505.06633. — [arXiv:2505.06633](https://arxiv.org/abs/2505.06633) find
  9. David R Hardoon, Sandor Szedmak, and John Shawe-Taylor. 2004. Canonical correlation analysis: An overview with application to learning methods. Neural computation, 16(12):2639–2664. find
  10. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531. find our card
  11. Nathaniel R. Robinson, Niyati Bafna, Xiluo He, Tom Lupicki, Lavanya Shankar, Cihan Xiao, Qi Sun, Kenton Murray, and David Yarowsky. 2025. find
  12. JHU IWSLT 2025 Low-resource System Description. In Proc. IWSLT. find
  13. SEAMLESS Communication Team. 2025. Joint speech and text machine translation for up to 100 languages. Nature, 637:587–593. find
  14. Claytone Sikasote, Eunice Mukonde, Md Mahfuz Ibn Alam, and Antonios Anastasopoulos. 2023. BIGC: A Multimodal Multi-Purpose Dataset for Bemba. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2062–2078. find
  15. Tiia Sildam, Andra Velve, and Tanel Alumäe. 2024. Finetuning end-to-end models for Estonian conversational spoken language translation. In Proceedings of the Seventh Workshop on Technologies for Machine Translation of Low-Resource Languages (LoResMT 2024), pages 166–174. find
  16. Ruiyan Sun and Satoshi Nakamura. 2026. Gradient-informed training for low-resource multilingual speech translation. arXiv preprint arXiv:2603.25836. — [arXiv:2603.25836](https://arxiv.org/abs/2603.25836) find
  17. Ziyu Yang, Guibin Chen, Yuxin Yang, Aoxiong Zeng, and Xiangquan Yang. 2026. Disentangling task conflicts in multi-task lora via orthogonal gradient projection. arXiv preprint arXiv:2601.09684. — [arXiv:2601.09684](https://arxiv.org/abs/2601.09684) find
  18. Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. 2020. Gradient surgery for multi-task learning. In Advances in Neural Information Processing Systems, volume 33, pages 5824–5836. find our card
  19. Shijie Zhu, Hui Zhao, Tianshu Wu, Pengjie Wang, Hongbo Deng, Jian Xu, and Bo Zheng. 2025. Gradient deconfliction via orthogonal projections onto subspaces for multi-task learning. In Proceedings of the 18th ACM International Conference on Web Search and Data Mining. find
  20. Harold Hotelling. 1936. Relations between two sets of variates. Biometrika, 28(3/4):321–377. find
  21. Chutong Meng and Antonios Anastasopoulos. 2025. GMU Systems for the IWSLT 2025 Low-Resource Speech Translation Shared Task. In Proc. IWSLT. find
  22. Maja Popović. 2015. chrf: character n-gram f-score for automatic mt evaluation. In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 392–395. find our card
  23. Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, and 1 others. 2024. Scaling speech technology to 1,000+ languages. Journal of Machine Learning Research, 25:1–52. find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/sun-2026-cuhksz-system-iwslt-2026-low/