Elamur Research⌕
☰
home / papers / 2026 / SpeechLM for Automatic Speech Recognition in Low-resource Languages

SpeechLM for Automatic Speech Recognition in Low-resource Languages

Paper · 2026 · Md Abdur Razzaq Riyadh, Eneko Agirre, Eva Navas et al.

Summary

The authors ask whether a general-purpose, instruction-tuned speech language model can be adapted into a supervised ASR system for two under-resourced languages, Basque and Maltese. They also ask which internal component matters most. They fine-tuned LLaMA-Omni, which chains a whisper speech encoder, a speech adaptor, LLaMA 3.1-8B-Instruct and a speech decoder, and ran extensive hyperparameter searches. They then ran three controlled experiments: plain fine-tuning; replacing the LLM with the Basque-adapted Latxa Instruct (called Latxa-Omni); and replacing the encoder with a whisper already fine-tuned on the target language. A 35-hour Basque subset served as a control for the effect of data size. With ample in-domain data, the SpeechLM beat a fine-tuned whisper-large-v3 on Basque. It suffered badly when data was scarce (Maltese, reduced Basque) and generalized worse to out-of-distribution Basque speech. The Basque LLM recovered much of that robustness. This is presented as the first systematic look at SpeechLM-based ASR in low-resource settings.

Key points

  • Average WER across test splits:

    • Basque: 8.09%
    • Maltese: 25.65%

  • Basque in-distribution: LLaMA-Omni improved on the fine-tuned whisper-large-v3 baseline by about 22%.

  • Maltese in-distribution: performance fell by roughly 50% relative to the baseline, which the authors attribute to the much smaller training set.

  • Out-of-distribution Basque (Faktoria), WER:

    ModelEUS-only portionOther portion
    whisper-large-v318.56%29.42%
    LLaMA-Omni25.23%54.81%
    Latxa-Omni21.87%31.54%
    LLaMA-Omni with adapted encoder22.88%45.14%

    Language-adapting the LLM helped more than language-adapting the encoder.

  • Maltese errors were mostly confusions between similar-sounding words and faulty word segmentation, often producing Italian-like forms. The authors read this as the LLM failing to learn Maltese grammar from limited data, even though the encoder captures the sounds.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of the Basque CountryLeioa · Spain57 works · 16,520 citations🎓University of MaltaMsida · Malta29 works · 894 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSignal Processing · 452 works

🧭 Topics 2Research subjects and languages the work is about.Speech & spoken language · 480 worksLow-resource languages · 267 works

Cite

Cite as Md Abdur Razzaq Riyadh, Eneko Agirre, Eva Navas and Claudia Borg. 2026. SpeechLM for Automatic Speech Recognition in Low-resource Languages. Proceedings of Speech Language Models in Low-Resource Settings: Performance, Evaluation, and Bias Analysis (SPEAKABLE) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2634e2pv97js

@inproceedings{riyadh2026speechlm,
  title = {{SpeechLM for Automatic Speech Recognition in Low-resource Languages}},
  author = {Md Abdur Razzaq Riyadh, Eneko Agirre, Eva Navas and Claudia Borg},
  booktitle = {Proceedings of Speech Language Models in Low-Resource Settings: Performance, Evaluation, and Bias Analysis (SPEAKABLE) @ LREC 2026},
  year = {2026},
  doi = {10.63317/2634e2pv97js},
  url = {https://doi.org/10.63317/2634e2pv97js}
}

References

11 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4218–4222, Marseille, France. European Language Resources Association. Julen Etxaniz, Oscar Sainz, Naiara Miguel, Itziar Aldabe, German Rigau, Eneko Agirre, Aitor Ormazabal, Mikel Artetxe, and Aitor Soroa. Latxa: An open language model and evaluation suite for basque. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14952–14972. Association for Computational Linguistics. ↗
  2. Qingkai Fang, Shoutao Guo, Yan Zhou, Zhengrui Ma, Shaolei Zhang, and Yang Feng. 2025. Llama-omni: Seamless speech interaction with large language models. In International Conference on Representation Learning, volume 2025, pages 57607–57624. ↗
  3. Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016. Deep Learning. MIT Press. http://www.deeplearningbook.org. Faktoria (EUS only) 18.56% 25.23% 21.87% 22.88% Faktoria (others) 29.42% 54.81% 31.54% 45.14% Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, et al. 2024. The Llama 3 Herd of Models. ↗
  4. Carlos Daniel Hernandez Mena, Albert Gatt, Andrea DeMarco, Claudia Borg, Lonneke van der Plas, Amanda Muscat, and Ian Padovani. 2020. MASRI-HEADSET: A Maltese corpus for speech recognition. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 6381–6388, Marseille, France. European Language Resources Association. ↗
  5. Oddur Kjartansson, Alexander Gutkin, Alena Butryna, Isin Demirsahin, and Clara Rivera. 2020. Open-Source High Quality Speech Datasets for Basque, Catalan and Galician. In Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced Languages (SLTU) and Collaboration and Computing for Under-Resourced Languages (CCURL), pages 21–27, Marseille, France. European Language Resources association. find
  6. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. find
  7. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust Speech Recognition via Large-Scale Weak Supervision. ↗
  8. Amparo Varona, Mikel Penagarikano, Germán Bordel, and Luis Javier Rodriguez-Fuentes. 2024. A Bilingual Basque–Spanish Dataset of Parliamentary Sessions for the Development and Evaluation of Speech Technology. Applied Sciences, 14(5):1951. ↗
  9. Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, and Jie Tang. 2024. GLM-4-Voice: Towards Intelligent and Human-Like End-to-End Spoken Chatbot. ↗
  10. Yuhao Zhang, Yuhao Du, Zhanchen Dai, Xiangnan Ma, Kaiqi Kou, Benyou Wang, and Haizhou Li. 2025a. Echox: Towards mitigating acoustic-semantic gap via echo training for speech-to-speech llms. ↗
  11. Yuhao Zhang, Zhiheng Liu, Fan Bu, Ruiyu Zhang, Benyou Wang, and Haizhou Li. 2025b. Soundwave: Less is More for Speech-Text Alignment in LLMs. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/riyadh-2026-speechlm-automatic-speech-recognition-low/