Elamur Research⌕
☰
home / papers / 2026 / Wise@DravidianLangTech 2026: Dialect-Aware Tamil Speech Classification and Recognition via Cross-Pipeline Embedding Transfer

Wise@DravidianLangTech 2026: Dialect-Aware Tamil Speech Classification and Recognition via Cross-Pipeline Embedding Transfer

Paper · 2026 · Ganesh Sundhar S, Hari Krishnan N, Gnanasabesan G et al.

Summary

The Wise team tackles two problems from the DravidianLangTech 2026 shared task on Tamil dialectal speech: identifying which of four dialect regions (Northern, Southern, Western, Central) an utterance comes from, and transcribing dialectal Tamil speech. The training data is small, about 9.22 hours across 5,134 clips, and the classes are unevenly distributed.

All audio first goes through a cleanup pipeline: picking the cleanest channel, removing DC offset, normalizing loudness to −23 LUFS, and applying neural denoising. For classification, the authors compare 12 setups. These combine two Wav2Vec2 backbones, five ways of pooling frames into an utterance vector, and frozen versus partially fine-tuned encoders. Their best classifier took first place by a wide margin.

For ASR, they reuse the classifier's pooled vector as a dialect embedding and feed it into Whisper-Tamil in one of two ways. It is either added at every encoder layer or supplied as an extra token for decoder cross-attention. The submitted ASR model was only partly trained and ranked 8th. The work matters because it shows how a dialect classifier can serve as a reusable conditioning signal for a low-resource Dravidian ASR system.

Key points

  • Preprocessing: channel selection by SNR, DC offset removal, ITU-R BS.1770 loudness normalization via pyloudnorm, and Facebook's Denoiser DNS64.
  • Classification setups: backbones were facebook/wav2vec2-large-xlsr-53 and Harveenchadha/vakyansh-wav2vec2-tamil-tam-250. The best setup used learned attentive pooling, unfroze the top 4 of 12 transformer layers, and trained with differential learning rates (10⁻⁵ for the encoder, 10⁻⁴ for the head).
  • Classification result: macro F1 of 0.79, ranked 1st. The runner-up, Wave2Word, scored 0.53.
  • ASR models: built on vasista22/whisper-tamil-small and whisper-tamil-medium. The residual-injection variant reached a WER of 0.90 (8th place), while the top team, CHMOD_777, reached 0.51.
  • Limitations and code: the authors note the small dataset, class imbalance (885 Central vs. 1,696 Northern clips), and noisy transcripts. Code is released on GitHub (Ganesh2609/DialectBasedSpeechProcessing).

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

🎓Amrita Vishwa VidyapeethamCoimbatore · India7 works · 31 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

Cite

Cite as Ganesh Sundhar S, Hari Krishnan N, Gnanasabesan G, Suriya KP and Jyothish Lal G. 2026. Wise@DravidianLangTech 2026: Dialect-Aware Tamil Speech Classification and Recognition via Cross-Pipeline Embedding Transfer. 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-1.71

@inproceedings{s2026wise,
  title = {{Wise@DravidianLangTech 2026: Dialect-Aware Tamil Speech Classification and Recognition via Cross-Pipeline Embedding Transfer}},
  author = {Ganesh Sundhar S, Hari Krishnan N, Gnanasabesan G, Suriya KP and Jyothish Lal G},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.71},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.71}
}

References

17 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Stephen E Levinson. 1986. Continuously variable duration hidden Markov models for automatic speech recognition. Computer Speech & Language, 1(1):29– 45. find
  2. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations. find
  3. Sanjeev Arora, Yingyu Liang, and Tengyu Ma. 2017. A simple but tough-to-beat baseline for sentence embeddings. In ICLR. find
  4. Koji Okabe, Takafumi Koshinaka, and Koichi Shinoda. 2018. Attentive statistics pooling for deep speaker embedding. In Interspeech. find
  5. Premjith B, Jyothish Lal G, Sowmya V, Bharathi Raja Chakravarthi, Rajeswari Natarajan, Nandhini K, Abirami Murugappan, Bharathi B, Kaushik M, Prasanth Sn, Aswin Raj R, and Vijai Simmon S. 2023. Findings of the shared task on multimodal abusive language detection and sentiment analysis in tamil and m… find
  6. Lawrence R Rabiner and Biing-Hwang Juang. 1993. Fundamentals of Speech Recognition. Prentice Hall. find
  7. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, volume 33, pages 12449–12460.
  8. 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
  9. 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
  10. Alexandre Defossez, Gabriel Synnaeve, and Yossi Adi. 2020. Real time speech enhancement in the waveform domain. arXiv preprint arXiv:2006.12847. 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. Proceedings of the 40th International Conference on Machine Learning.
  12. David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. 1986. Learning representations by backpropagating errors. Nature, 323:533–536. find
  13. Ganesh Sundhar S, Durai Singh K, Gnanasabesan G, Hari Krishnan N, and Mc Dhanush. 2025a. Wise@LT-EDI-2025: Combining classical and neural representations with multi-scale ensemble learning for code-mixed hate speech detection. In Proceedings of the 5th Conference on Language, Data and Knowledge: Fif… find
  14. Ganesh Sundhar S, Hari Krishnan N, Arun Prasad T D, Shruthikaa V, and Jyothish Lal G. 2025b. CrewX@LT-EDI-2025: Transformer-based Tamil ASR fine-tuning with AVMD denoising and GRU-VAD for enhanced transcription accuracy. In Proceedings of the 5th Conference on Language, Data and Knowledge: Fifth Wor… find
  15. K Sreelakshmi, B Premjith, Bharathi Raja Chakravarthi, and K P Soman. 2024. Detection of hate speech and offensive language codemix text in dravidian languages using cost-sensitive learning approach. IEEE Access. find
  16. C. J. Van Rijsbergen. 1979. Information Retrieval. Butterworth-Heinemann. find
  17. Felix Weninger, Hakan Erdogan, Shinji Watanabe, and 1 others. 2015. Speech enhancement with LSTM recurrent neural networks and its application to noise-robust ASR. In International Conference on Latent Variable Analysis and Signal Separation. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/s-2026-wise-dravidianlangtech-2026-dialect-aware/