Elamur Research⌕
☰
home / papers / 2026 / CATENG Submission for the IWSLT 2026: Dialectal and Low-resource Speech Translation Task

CATENG Submission for the IWSLT 2026: Dialectal and Low-resource Speech Translation Task

Summary

This paper describes the CATENG team's entries to the IWSLT 2026 shared task on dialectal and low-resource speech translation, which covers translating spoken Catalan into English text. Catalan has good text resources but is still underserved in speech technology. Its Eastern and Western varieties, along with mixing with Spanish, make recognition harder. The team submitted three unconstrained systems, trained on the roughly 15-hour IWSLT 2026 Catalan dataset plus external parallel text.

Two of the systems are cascades that feed a fine-tuned NLLB-200 translation model. The first uses a Mamba-based recognizer (ConMamba, the primary run). The second uses a Catalan-adapted Whisper-v3, further tuned on the task data. The third system is a direct SpeechT5 model trained with audio augmentation.

The Whisper-v3 cascade scored highest, and both pipelines beat the end-to-end model. Because the cascades share the same MT component, the authors conclude that recognition quality, rather than translation, is the limiting factor. The Mamba-based recognizer was already close to Whisper-v3. This edition is also the first in which Catalan appears as a source language in the track.

Key points

  • Official scores: Whisper-v3 + NLLB reached 44.7 BLEU / 65.1 chrF, ConMamba + NLLB reached 43.2 / 64.3, and SpeechT5 (end-to-end) reached 41.3 / 63.1. The gap between best and worst is 3.4 BLEU.
  • The MT component is NLLB-200 (1.3B), fine-tuned for CA–EN with these settings:
    • sequences up to 128 tokens
    • 10 epochs, batch size 8
    • beam search with 5 beams
    • fixed seed 65
  • The Whisper run starts from the Projecte AINA checkpoint faster-whisper-large-v3-ca-3catparla. It is then fine-tuned on the task's Catalan transcriptions as ASR targets.
  • SpeechT5 has a 12-layer encoder and a 6-layer decoder and was pretrained on 960 h of LibriSpeech. It was augmented with nlpaug (noise, temporal distortion, signal perturbation), which doubled the training audio.
  • The dataset is public on GitHub (rjzevallos/IWSLT_2026_Catalan_Dataset). The constrained setting no longer exists in this track.

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-1.18

@inproceedings{zevallos2026cateng,
  title = {{CATENG Submission for the IWSLT 2026: Dialectal and Low-resource Speech Translation Task}},
  author = {Zevallos, Rodolfo Joel and Casals, Marc and Ortega, John E. and Carraro, Fabrício and Buitrago, Pol and Cámbara, Guillermo},
  booktitle = {Proceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.iwslt-1.18},
  url = {https://doi.org/10.18653/v1/2026.iwslt-1.18}
}

References

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

  1. Željko Agić and Ivan Vulić. 2019. JW300: A wide-coverage parallel corpus for low-resource languages. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3204– 3210, Florence, Italy. Association for Computational Linguistics. \#\# Conclusion This paper su… find
  2. The work reinforces the comparison between end-to-end and cascaded systems. In particular, the system based on Whisper-v3 ASR and NLLB MT achieved the best performance, highlighting the importance of strong multilingual pre-trained models. In comparison, the other cascaded system that used ConMamba… find
  3. Junyi Ao, Rui Wang, Long Zhou, Chengyi Wang, Shuo Ren, Yu Wu, Shujie Liu, Tom Ko, Qing Li, Yu Zhang, Zhihua Wei, Yao Qian, Jinyu Li, and Furu Wei. 2022. SpeechT5: Unified-modal encoder-decoder pre-training for spoken language processing. In Proceedings of the 60th Annual Meeting of the Association f… find
  4. Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M. Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. arXiv preprint arXiv:1912.06670. find our card
  5. Carme Armentano-Oller, Montserrat Marimon, and Marta Villegas. 2024. Becoming a high-resource language in speech: The Catalan case in the Common Voice corpus. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). find
  6. Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2021. XLS-R: Self-supervised cross-lingual speech representation learning at scale. arXiv preprint arXi… find our card
  7. Barcelona Supercomputing Center, Language Technologies Unit. 2025. SalamandraTA: A translation-specialised variant of salamandra. Model release, Hugging Face BSC-LT/salamandrata-7b. find
  8. William Chen, Brian Yan, Jiatong Shi, Yifan Peng, Soumi Maiti, and Shinji Watanabe. 2023. Improving massively multilingual ASR with auxiliary CTC objectives. In Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). find
  9. 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 Proc. IEEE Spoken Language Technology Workshop (SLT). find our card
  10. Projecte AINA. 2024. Common Voice benchmark catalan accents. Derived from Catalan Common Voice v17, stratified by accent (Balearic, Central, Northern, Northwestern, Valencian) and gender. find
  11. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning (ICML). find our card
  12. Aitor Gonzalez-Agirre, Marc Pàmies, Joan Llop, Irene Baucells, Severino Da Dalt, Daniel Tamayo, José Javier Saiz, Ferran Espuña, Jaume Prats, Javier Aula-Blasco, Mario Mina, Adrián Rubio, Alexander Shvets, Anna Sallés, Iñaki Lacunza, Iñigo Pikabea, Jorge Palomar, Júlia Falcão, Lucía Tormo, and 4 oth… — [arXiv:2502.08489](https://arxiv.org/abs/2502.08489) find
  13. Baybars Külebi, Carme Armentano-Oller, Carlos Rodriguez-Penagos, and Marta Villegas. 2022. ParlamentParla: A speech corpus of Catalan parliamentary sessions. In Proceedings of the Workshop ParlaCLARIN III within the 13th Language Resources and Evaluation Conference, pages 125–130, Marseille, France.… find
  14. Changhan Wang, Anne Wu, and Juan Pino. 2021. CoVoST 2 and massively multilingual speech translation. In Proc. Interspeech. find
  15. Rodolfo Zevallos, Núria Bel, Guillermo Cámbara, Mireia Farrús, and Jordi Luque. 2022. Data Augmentation for Low-Resource Quechua ASR Improvement. In Interspeech 2022, pages 3518–3522. find
  16. Rodolfo Joel Zevallos Salazar, Martí Cortada García, Sarah Solito, Carlos Daniel Hernández Mena, Alexandre Peiró Lilja, and Francisco Javier Hernando Pericás. 2025. Assessing the performance and efficiency of mamba asr in low-resource scenarios. In Interspeech 2025: Rotterdam, The Netherlands, 17-21… find
  17. Edward Ma. 2019. Nlp augmentation. https://github.com/makcedward/nlpaug. ↗
  18. NLLB Team, 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, and 20 o… find
  19. Alex Peiró-Lilja, Carme Armentano-Oller, José Giraldo, Wendy Elvira-García, Ignasi Esquerra, Rodolfo Zevallos, Cristina España-Bonet, Martí Llopart-Font, Baybars Külebi, and Mireia Farrús. 2026. Lafrescat: A studio-quality catalan multi-accent speech dataset for text-to-speech synthesis. Computer Sp… find
  20. Alex Peiró-Lilja, Martí Llopart-Font, Carme Armentano-Oller, José Giraldo, and Ignasi Esquerra. 2024. LaFresCat: A Catalan multi-accent speech dataset for text-to-speech. In Proc. find
  21. Yifan Peng, Jinchuan Tian, Brian Yan, Dan Berrebbi, Xuankai Chang, Xinjian Li, Jiatong Shi, Siddhant Arora, William Chen, Roshan Sharma, Wangyou Zhang, Yui Sudo, Muhammad Shakeel, Jee-weon Jung, Soumi Maiti, and Shinji Watanabe. 2023. Reproducing Whisper-style training using an open-source toolkit a… find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/zevallos-2026-cateng-submission-iwslt-2026-dialectal/