Elamur Research⌕
☰
home / papers / 2026 / Mapudungun-Spanish Speech Translation: A Low-Resource End-to-End System for the IWSLT 2026 Shared Task

Mapudungun-Spanish Speech Translation: A Low-Resource End-to-End System for the IWSLT 2026 Shared Task

Summary

The authors built an end-to-end speech-to-text translation system from Mapudungun (arn), an endangered Indigenous language of Chile and Argentina, into Spanish. It was their entry for Task 1 of the IWSLT 2026 Low-Resource Speech Translation shared task. They used the provided corpus by Duan et al. (2020), about 142 hours of healthcare-themed interviews from the AVENUE project, and adapted NVIDIA's Canary-1B-v2 (FastConformer encoder, Transformer decoder, NeMo framework). The backbone stayed frozen, and only a small linear adapter in the encoder was trained. Canary translates only to and from English, so they used the eng → spa direction and let the "English" source tag stand in for Mapudungun. This relies on the model's existing acoustic knowledge and its ability to generate Spanish. Three runs compared on-the-fly audio augmentation, a plain setup, and filtering out utterances longer than 15 seconds. The filtered run worked best. Absolute scores were very low, and outputs sometimes degenerated into repetitive Spanish phrases. The work shows that cheap adaptation of multilingual foundation models to Indigenous languages is possible but still very limited.

Key points

  • Data: The Mapudungun corpus contains 142 h of speech, cleaned from 170 h recorded in 2001–2005. It covers three dialects: Nguluche (~110 h), Lafkenche (~20 h) and Pewenche (~10 h). It uses a 28-grapheme supra-dialectal orthography rather than the Azumchefe standard.
  • Method: PEFT with a linear adapter (1024 → 8, Swish activation) reduces trainable parameters from a reported 1.7B to 589k. Other settings were bfloat16 training, Lhotse duration bucketing and lower-cased Spanish without punctuation. The hardware was modest, with GPUs of 24–32 GB.
  • Runs: The three runs were Augmented (Main), Vanilla and Filtered (audio ≤15 s). On the development set, the post-submission Filtered* version reached BLEU 3.7 and WER 1.56. The other two runs scored BLEU ~1.0–1.05 and WER 2.3–2.5.
  • Official test set: Filtered scored BLEU 0.82 and chrF2++ 14.31. Main scored BLEU 0.73 and chrF2++ 12.70, and Vanilla scored BLEU 0.34 and chrF2++ 6.92.
  • Observations:
    • Augmentation helped little, probably because the corpus is already noisy.
    • The model captured some region-specific words such as huinca and mapuche.
    • Two-GPU training was unstable.
    • The authors plan to extend the approach to Quechua and Nahuatl.
    • Code: github.com/umoqnier/iwstl-low-resource-experiments.

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

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.iwslt-1.26

@inproceedings{martinez2026mapudungun,
  title = {{Mapudungun-Spanish Speech Translation: A Low-Resource End-to-End System for the IWSLT 2026 Shared Task}},
  author = {Martínez, Diego Alberto Barriga and Gazque, Amilkar and Elizalde, Mikel Segura and Mena, Carlos Daniel Hernandez and Gutierrez-Vasques, Ximena and Ruiz, Ivan Vladimir Meza},
  booktitle = {Proceedings of the 23rd International Conference on Spoken Language Translation (IWSLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.iwslt-1.26},
  url = {https://doi.org/10.18653/v1/2026.iwslt-1.26}
}

References

12 references, 1 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. Mingjun Duan, Carlos Fasola, Sai Krishna Rallabandi, Rodolfo Vega, Antonios Anastasopoulos, Lori Levin, and Alan W Black. 2020. A resource for computational experiments on Mapudungun. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 2872–2877, Marseille, France. Euro… find
  3. Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and 1 others. 2020. Conformer: Convolution-augmented transformer for speech recognition. arXiv preprint arXiv:2005.08100. find our card
  4. Harald Hammarström, Robert Forkel, Martin Haspelmath, and Sebastian Bank. 2024. Glottolog 5.2 -mapudungun. Accessed 2026-04-21. find
  5. Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790–2799. PMLR. find
  6. Ivér Jordal. 2023. Audiomentations: A python library for audio data augmentation. https://iver56.github.io/audiomentations/. Accessed: 2026-04-23. ↗
  7. Oleksii Kuchaiev, Jason Li, Huyen Nguyen, Oleksii Hrinchuk, Ryan Leary, Boris Ginsburg, Samuel Kriman, Stanislav Beliaev, Vitaly Lavrukhin, Jack Cook, and 1 others. 2019. Nemo: a toolkit for building ai applications using neural modules. arXiv preprint arXiv:1909.09577. find
  8. Lorraine Levin, Rodolfo M Vega, Jaime G Carbonell, Ralf D Brown, Alon Lavie, Eliseo Cañulef, and Carolina Huenchullan. 2000. Data collection and language technologies for mapudungun. find
  9. NVIDIA. 2025. Canary-1b-v2: Multitask speech transcription and translation model. https://huggingface.co/nvidia/canary-1b-v2. Hugging Face model card. ↗
  10. Dima Rekesh, Nithin Rao Koluguri, Samuel Kriman, Somshubra Majumdar, Vahid Noroozi, He Huang, Oleksii Hrinchuk, Krishna Puvvada, Ankur Kumar, Jagadeesh Balam, and Boris Ginsburg. 2023. Fast conformer with linearly scalable attention for efficient speech recognition. Preprint, arXiv:2305.05084. — [arXiv:2305.05084](https://arxiv.org/abs/2305.05084) find
  11. Monica Sekoyan, Nithin Rao Koluguri, Nune Tade-vosyan, Piotr Zelasko, Travis Bartley, Nikolay Karpov, Jagadeesh Balam, and Boris Ginsburg. 2025. Canary-1b-v2 & parakeet-tdt-0.6b-v3: Efficient and high-performance models for multilingual asr and ast. Preprint, arXiv:2509.14128. — [arXiv:2509.14128](https://arxiv.org/abs/2509.14128) find
  12. Piotr Żelasko, Dominika Gajewska, and David Povey. 2023. Lhotse: a python toolkit for working with audio data for speech applications. In Proceedings of the 2022 IEEE Spoken Language Technology Workshop (SLT), pages 129–136. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/martinez-2026-mapudungun-spanish-speech-translation-low/