Elamur Research⌕
☰
home / papers / 2026 / CUET_InferX@DravidianLangTech 2026: Shared Task on Dialect Based Speech Recognition and Classification in Tamil

CUET_InferX@DravidianLangTech 2026: Shared Task on Dialect Based Speech Recognition and Classification in Tamil

Paper · 2026 · Md. Ashraful Islam Semon, Jihadul Islam, Ratnajit Dhar et al.

Summary

The CUET_InferX team describes its systems for the DravidianLangTech@ACL 2026 shared task on Tamil dialects. The task has two parts: transcribing spontaneous dialectal Tamil speech (ASR) and assigning each recording to one of four regions: Northern, Southern, Western or Central. The data come from the multi-dialect Tamil corpus by Bharathi et al. (2025). For ASR, the authors fine-tuned Whisper-large-v3 with LoRA and trained a CTC-based Wav2Vec2 XLSR-53 model whose decoding uses a KenLM 4-gram language model. They merged the two outputs by aligning word sequences and settling disagreements with language-model scores. This combination reached a WER of 0.5410, which placed second. For dialect classification, the team ran a weighted soft-voting ensemble of text transformers over the ASR transcripts. It scored a macro F1 of only 0.2209, placing ninth. Error analysis showed that KenLM tended to replace rare dialect words with standard forms. A later fusion of audio and text features roughly doubled F1. The authors conclude that acoustic cues are essential for identifying Tamil dialects.

Key points

  • Data split: training has 5,134 utterances (Southern 1427, Northern 1696, Western 1126, Central 885); test has 579 (212/189/116/62).
  • ASR setup:
    • Whisper-large-v3 was fine-tuned through Unsloth with LoRA (r = 64, α = 64) on q_proj/v_proj.
    • XLSR-53 was trained with CTC.
    • KenLM decoding used LM weight 0.4, word bonus 1.5 and beam width 500.
    • Outputs were merged via SequenceMatcher alignment, with a slight preference for XLSR.
  • ASR results (WER on test): XLSR 0.6127, Whisper 0.5510, XLSR+KenLM 0.5491, final ensemble 0.5410 (CER 0.1690).
  • Dialect classification:
    • The ensemble combined XLM-RoBERTa (weight 1.4), TamilBERT (1.1), IndicBERT and MuRIL (0.9 each).
    • Transcripts from Whisper alone gave better classification than the ensemble's transcripts, because Whisper kept more colloquial forms.
  • Post-competition fusion: combining Whisper acoustic embeddings with XLM-RoBERTa text embeddings at an early stage raised macro F1 to about 0.40.

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

🎓Chittagong University of Engineering & TechnologyChittagong · Bangladesh25 works · 29 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 1Research subjects and languages the work is about.Speech & spoken language · 480 works

Cite

Cite as Md. Ashraful Islam Semon, Jihadul Islam, Ratnajit Dhar and Hasan Murad. 2026. CUET_InferX@DravidianLangTech 2026: Shared Task on Dialect Based Speech Recognition and Classification in Tamil. Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.dravidianlangtech-1.27

@inproceedings{semon2026cuet,
  title = {{CUET\_InferX@DravidianLangTech 2026: Shared Task on Dialect Based Speech Recognition and Classification in Tamil}},
  author = {Md. Ashraful Islam Semon, Jihadul Islam, Ratnajit Dhar and Hasan Murad},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.27},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.27}
}

References

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

  1. Priyobroto Acharya, Soham Chaudhuri, Sayan Das, Dipanjan Saha, and Dipankar Das. 2025. Junlp@lt-edi-2025: Efficient low-rank adaptation of whisper for inclusive tamil speech recognition targeting vulnerable populations. In Proceedings of the 5th Conference on Language, Data and Knowledge: Fifth Work… find
  2. A Akhilesh, Brinda P, Keerthana S, Deepa Gupta, and Susmitha Vekkot. 2022. Tamil speech recognition using xlsr wav2vec2.0 & ctc algorithm. In 2022 13th International Conference on Computing Communication and Networking Technologies (ICCCNT), pages 1–6. IEEE. find
  3. B. Bharathi, Bharathi Raja Chakravarthi, Shunmuga Priya Muthusamy Chinnan, S. Saranya, and S. Suhasini. 2026. Findings in Tamil Dialect Speech Recognition and Classification. In Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages. Association for Co… find
  4. B Bharathi, S Saranya, P Vijayalakshmi, and T Nagarajan. 2025. Multi-dialect speech corpus creation for enhancing tamil automatic speech recognition. Circuits, Systems, and Signal Processing, pages 1–19. find
  5. Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, and Michael Auli. 2020a. Unsupervised cross-lingual representation learning for speech recognition. arXiv preprint arXiv:2006.13979. find
  6. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Unsupervised cross-lingual representation learning at scale. In Proceedings of ACL. find
  7. Michael Han Daniel Han and Unsloth team. 2023. Unsloth. find
  8. Kenneth Heafield. 2011. Kenlm: Faster and smaller language model queries. In Proceedings of the sixth workshop on statistical machine translation, pages 187–197.
  9. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685.
  10. Raviraj Joshi. 2022. L3cube-hindbert and devbert: Pre-trained bert transformer models for devanagari based hindi and marathi languages. arXiv preprint arXiv:2211.11418. — [arXiv:2211.11418](https://arxiv.org/abs/2211.11418) find
  11. Divyanshu Kakwani, Anoop Kunchukuttan, Satish Golla, Gokul N.C., Avik Bhattacharyya, Mitesh M. Khapra, and Pratyush Kumar. 2020. IndicNLPSuite: Monolingual Corpora, Evaluation Benchmarks and Pre-trained Multilingual Language Models for Indian Languages. In Findings of EMNLP.
  12. Simran Khanuja and 1 others. 2021. Muril: Multilingual representations for indian languages. arXiv preprint arXiv:2103.10730.
  13. Zoey Liu, Nitin Venkateswaran, Eric Le Ferrand, and Emily Prud’hommeaux. 2024. How important is a language model for low-resource asr? In Findings of the Association for Computational Linguistics: ACL 2024, pages 206–213, Bangkok, Thailand. Association for Computational Linguistics. find
  14. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning, pages 28448–28481. PMLR.
  15. S. Saranya, B. Bharathi, S. Gomathy Dhanya, and Aishwarya Krishnakumar. 2024. Real-time continuous tamil dialect speech recognition and summarization. Circuits, Systems, and Signal Processing, 44(4):2855– 2881. find
  16. Dhanya Srinivasan, B. Bharathi, Thenmozhi Durairaj, and B. Senthil Kumar. 2022. Ssncse_nlp@lt-edi-acl2022: Speech recognition for vulnerable individuals in tamil using pre-trained xlsr models. In Proceedings of the Second Workshop on Language Technology for Equality, Diversity and Inclusion, pages 3… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/semon-2026-cuet-inferx-dravidianlangtech-2026-shared/