Elamur Research⌕
☰
home / papers / 2026 / Konkani Daan: A Community-Driven Culturally Grounded Speech Corpus for Low-Resource ASR

Konkani Daan: A Community-Driven Culturally Grounded Speech Corpus for Low-Resource ASR

Paper · 2026 · Milind Shivolkar, Vaibhav Gawas and Jyoti Pawar

Summary

Multilingual ASR systems that officially support Konkani can still fail on culturally dense speech, such as Goan place and temple names, administrative and devotional vocabulary, Portuguese-influenced words and mixed numeric forms. To address this gap, the authors built Konkani Daan at Goa University's Vidyapati Lab. It is a web platform where native speakers record randomly assigned text prompts on their own devices. Prompts come from Konkani WordNet example sentences and AI4Bharat's DMU dataset. Speakers report their home region within Goa, and the interface checks ambient noise before recording. The Devanagari transcripts keep community spelling habits rather than applying heavy normalisation. In zero-shot mode, AI4Bharat IndicConformer-600M scored 46.46% WER and 15.47% CER on the development set, which shows a sizeable domain and cultural gap. The authors also analysed errors from this model and from an earlier Konkani-trained Wav2Vec2 system. Many errors came from compound splitting, digits versus number words, distorted named entities and spelling variants, not only from failures to recognise the speech itself. The work argues for culturally informed corpora and normalisation-aware evaluation for low-resource Indian languages.

Key points

  • Corpus size and format: Konkani Daan holds over 43.9 hours of read (prompted) Goan Konkani speech, recorded as 16 kHz WAV with Devanagari transcripts and speaker-region metadata.
  • Contributors: 42 people have joined the platform, but only 18 are active (more than 50 recordings each). Contributions range from 55 to 1450 recordings per active person.
  • Prompt sources: Konkani WordNet supplies 32,804 sentences and the AI4Bharat DMU dataset supplies 31,894 sentences.
  • Baseline results: IndicConformer-600M with CTC decoding, run zero-shot on 379 development utterances, reached 46.46% WER and 15.47% CER. A previously developed Konkani-trained Wav2Vec2 model was additionally evaluated in a cross-corpus setting.
  • Access: The data is not openly downloadable. Researchers can request access for non-commercial use, subject to review by the host institution. The corpus will be released gradually as validation and ethical review continue.

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

🎓Goa UniversityTaleigao · India3 works · 4 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 Milind Shivolkar, Vaibhav Gawas and Jyoti Pawar. 2026. Konkani Daan: A Community-Driven Culturally Grounded Speech Corpus for Low-Resource ASR. Proceedings of the 8th Workshop on Indian Language Data: Resources and Evaluation.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2auy49fgeg5o

@inproceedings{shivolkar2026konkani,
  title = {{Konkani Daan: A Community-Driven Culturally Grounded Speech Corpus for Low-Resource ASR}},
  author = {Milind Shivolkar, Vaibhav Gawas and Jyoti Pawar},
  booktitle = {Proceedings of the 8th Workshop on Indian Language Data: Resources and Evaluation},
  year = {2026},
  doi = {10.63317/2auy49fgeg5o},
  url = {https://doi.org/10.63317/2auy49fgeg5o}
}

References

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

  1. Andrew Morris, Viktoria Maier, and Phil Green. 2004. From wer and ril to mer and wil: Improved evaluation measures for connected speech recognition. In Proceedings of Interspeech. find
  2. AI4Bharat. 2022. Digital multilingual utterances (dmu) dataset. https://ai4bharat.org. ↗
  3. Pratyush Anand, Ashish Gupta, Anshuman Goyal, Anoop Kunchukuttan, Animesh Prasad, et al. 2023. Indicconformer: A multilingual speech recognition model for indian languages. In Proceedings of Interspeech. find
  4. Manisha Prabhugaonkar, Prashant Bhandari, and Sangeeta Kamat. 2017. Building the konkani wordnet. In Proceedings of the Global WordNet Conference. find
  5. Alexei Baevski, Henry 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 (NeurIPS).
  6. Animesh Prasad et al. 2023. Shrutilipi: Building speech recognition datasets for indian languages. In Proceedings of the International Conference on Speech and Computer (SPECOM). find
  7. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision.
  8. Emily M. Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling bet-Anton Ragni, Mark JF Gales, Sibsankar Rath, and Yongqi Wang. 2014. Data augmentation for low resource languages. Interspeech. find
  9. Milind Shivolkar et al. 2026. Cross-domain generalisation in low-resource asr: A multi-corpus investigation of konkani using wav2vec2. In Proceedings of the International Conference on Intelligent Computing and Communication (ICICI). Accepted. find
  10. Anirudh Srinivasan et al. 2023. Indicvoices: A large-scale multilingual speech dataset for indian languages. In Proceedings of Interspeech. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shivolkar-2026-konkani-daan-community-driven-culturally/