Elamur Research⌕
☰
home / papers / 2026 / Nwāchā Munā: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR

Nwāchā Munā: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR

Paper · 2026 · Rishikesh Kumar Sharma, Safal Narshing Shrestha, Jenny Poudel et al.

Summary

Nepal Bhasha (Newari) is a Tibeto-Burman language of the Kathmandu Valley that UNESCO lists as endangered. Almost no transcribed speech exists for it, and the one earlier ASR system scored its output in Romanized transliteration rather than native Devanagari. The authors built Nwāchā Munā, a 5.39-hour corpus transcribed by hand in Devanagari. It combines smartphone recordings from 18 native speakers with roughly an hour of existing web audio, which community members converted from Roman script to Devanagari. The text prompts came from Wikipedia, OSCAR, newspapers, Bloom Library materials, school textbooks and hand-written conversational sentences, and were normalized (for example, numerals written out as words).

On this corpus they set up the first script-preserving benchmark. It tests whether transfer from a closely related neighbouring language (Nepali → Newari) can compete with large multilingual pretraining. Fine-tuning a Nepali Conformer with data augmentation brought CER down to about the level of Whisper-Small while using far fewer parameters. This makes proximal transfer an inexpensive route for other low-resource languages in the region. The data and benchmarks are openly released.

Key points

  • Corpus scale: Nwāchā Munā has 5,727 utterances, 27,644 words and 8,599 unique words, mostly short sentences (4,323 of 0–5 words). Audio is 16 kHz mono WAV.
  • Audio sources: 4 h 21 min was field-recorded by 18 speakers from Banepa, Dhulikhel, Panauti and Patan, balanced by age and gender. About 1 h was added from Meelen et al. (2024) audio, re-transcribed into Devanagari.
  • Main result: Fine-tuning a Nepali Conformer cut CER from 52.54% (zero-shot) to 17.59% with augmentation, roughly equal to Whisper-Small.
  • Decoding: Shallow fusion with an external KenLM n-gram language model was also tested.
  • Comparison with prior work: Earlier Newari ASR by Meelen et al. reported 12% CER, but scored in Romanized form. This benchmark is the first to evaluate output in Devanagari script.

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

🎓Kathmandu UniversityKathmandu · Nepal7 works · 20 citations

 Fields, topics and concepts

📚 Shelves 7The 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 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 worksSignal Processing · 452 works

🧭 Topics 2Research subjects and languages the work is about.Evaluation & benchmarks · 1,220 worksSpeech & spoken language · 480 works

Cite

Cite as Rishikesh Kumar Sharma, Safal Narshing Shrestha, Jenny Poudel, Rupak Tiwari, Arju Shrestha, Rupak Raj Ghimire and Bal Krishna Bal. 2026. Nwāchā Munā: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR. Proceedings of the Second workshop on Challenges in Processing South Asian Languages (CHiPSAL2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3njkgpwzsn7k

@inproceedings{sharma2026nwacha,
  title = {{Nwāchā Munā: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR}},
  author = {Rishikesh Kumar Sharma, Safal Narshing Shrestha, Jenny Poudel, Rupak Tiwari, Arju Shrestha, Rupak Raj Ghimire and Bal Krishna Bal},
  booktitle = {Proceedings of the Second workshop on Challenges in Processing South Asian Languages (CHiPSAL2026)},
  year = {2026},
  doi = {10.63317/3njkgpwzsn7k},
  url = {https://doi.org/10.63317/3njkgpwzsn7k}
}

References

28 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shweta Bansal, Shweta Sinha, and Shyam S. Agrawal. 2020. Acoustic-phonetic approach for ASR of less resourced languages using monolingual and cross lingual information. In Proceedings of the 1st Joint SLTU and CCURL Workshop (SLTU-CCURL 2020), pages 167– 171. European Language Resources Association… find
  2. Bharat Bhatta, Basanta Joshi, and Ram Krishna Maharjhan. 2020. Nepali speech recognition using CNN, GRU and CTC. In Proceedings of the 32nd Conference on Computational Linguistics and Speech Processing (ROCLING 2020), pages 238–246. The Association for Computational Linguistics and Chinese Language… find
  3. Yao-Fei Cheng, Li-Wei Chen, Hung-Shin Lee, and Hsin-Min Wang. 2024. Exploring the impact of data quantity on asr in extremely low-resource languages. find
  4. Rupak Raj Ghimire, Bal Krishna Bal, and Prakash Poudyal. 2023a. Active learning approach for fine-tuning pre-trained ASR model for a low-resourced language: A case study of Nepali. In Proceedings of the 20th International Conference on Natural Language Processing (ICON), pages 82–89, Goa University,… find
  5. Rupak Raj Ghimire, Bal Krishna Bal, Balaram Prasain, and Prakash Poudyal. 2023b. Pronunciation-aware syllable tokenizer for Nepali automatic speech recognition system. In Proceedings of the 20th International Conference on Natural Language Processing (ICON), pages 36–43, Goa University, Goa, India.… find
  6. Rupak Raj Ghimire, Prakash Poudyal, and Bal Krishna Bal. 2025. Improving accuracy of low-resource asr using rule-based character constituency loss (rbccl). In Proceedings of the First Workshop on Challenges in Processing South Asian Languages (CHiPSAL), pages 61–70. Association for Computational Lin… find
  7. Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. 2020. Conformer: Convolution-augmented transformer for speech recognition. In Proceedings of Interspeech 2020, pages 5036–5040. ISCA.
  8. Kenneth Heafield. 2011. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland. Association for Computational Linguistics.
  9. Thennal D K, Jesin James, Deepa Padmini Gopinath, and Muhammed Ashraf K. 2025. Advocating character error rate for multilingual ASR evaluation. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 4941–4950, Albuquerque, New Mexico. Association for Computational Linguistic… find
  10. Tahir Javed, Sumanth Doddapaneni, Abhigyan Raman, Kaushal Santosh Bhogale, Gowtham Ramesh, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh M. Khapra. 2022. Towards building ASR systems for the next billion users. In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence (AAAI-22). find
  11. Manish Dhakal, Arman Chhetri, Aman Kumar Gupta, Prabin Lamichhane, Suraj Pandey, and Automatic speech Subarna Shakya. 2022. recognition for the nepali language using CNN, bidirectional LSTM and ResNet. In 2022 International Conference on Inventive Computation Technologies (ICICT), pages 515–521. IEE… find
  12. Basanta Joshi and Rupesh Shrestha. 2023. Nepali speech recognition using self-attention networks. International Journal of Innovative Computing, Information and Control, 19(6):1769– 1784. find
  13. David M. Eberhard, Gary F. Simons, and Charles D. Fennig. 2026. Ethnologue: Languages of the World, twenty-ninth edition. SIL International, Dallas, Texas. Online version: http://www.ethnologue.com. ↗
  14. Marieke Meelen, Alexander O’Neill, and Rolando Coto-Solano. 2024. End-to-end speech recognition for endangered languages of nepal. In Proceedings of the 7th Workshop on Computational Methods for Endangered Languages (ComputEL-7), pages 83–93. Association for Computational Linguistics. find
  15. Christopher Moseley, editor. 2010. Atlas of the World’s Languages in Danger, 3rd edition. UN-ESCO Publishing, Paris. find
  16. Pedro Javier Ortiz Suárez, Laurent Romary, and Benoît Sagot. 2020. A monolingual approach to contextualized word embeddings for mid-resource languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1703–1714, Online. Association for Computational L…
  17. Kiran Pantha, Rupak Raj Ghimire, and Bal Krishna Bal. 2025. Speech personalization using parameter efficient fine-tuning for Nepali speakers. In Proceedings of the 5th Conference on Language, Data and Knowledge: Fifth Workshop on Language Technology for Equality, Diversity, Inclusion, pages 190–199,… find
  18. Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Specaugment: A sim-Quoc V. Le. 2019. ple data augmentation method for automatic speech recognition. In Proceedings of Interspeech 2019, pages 2613–2617. ISCA. find
  19. Silchar, Silchar, India. NLP Association of India (NLPAI). find
  20. Manish K. Ssarma, Avaas Gajurel, Anup Pokhrel, and Basanta Joshi. 2017. Hmm based isolated word nepali speech recognition. In 2017 International Conference on Machine Learning and Cybernetics (ICMLC), volume 1, pages 1–6. IEEE. find
  21. Shubham Toshniwal, Anjuli Kannan, Chung-Cheng Chiu, Yonghui Wu, Tara N Sainath, and Karen Livescu. 2018. A comparison of techniques for language model integration in encoder-decoder speech recognition. In 2018 IEEE Spoken Language Technology Workshop (SLT), pages 369–375. IEEE. find
  22. Ashish Vaswani, Llion Jones, Noam Shazeer, Niki Parmar, Aidan N. Gomez, Jakob Uszkoreit, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30.
  23. Shishir Paudel, Bal Krishna Bal, and Dhiraj Shrestha. 2023. Large vocabulary continous speech recognition for Nepali language using CNN and transformer. In Proceedings of the 4th Conference on Language, Data and Knowledge, pages 328–333, Vienna, Austria. NOVA CLUNL, Portugal. find
  24. Post Report. 2024. Nepali, tamang, nepal bhasha official languages of bagmati. The Kathmandu Post. Accessed: 2026-02-26. find
  25. Jenny Poudel, Ankit Dahal, Rishikesh Kumar Sharma, Rupak Tiwari, Rupak Raj Ghimire, and Bal Krishna Bal. 2025. Nepconformer: A conformer-based nepali automatic speech recognition system. In International Conference on Computing and Machine Learning, pages 167–178. Springer. find
  26. Lawrence R. Rabiner. 1989. A tutorial on hidden markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2):257–286. find
  27. P Regmi, A Dahal, and B Joshi. 2019. Nepali speech recognition using rnn-ctc model. International Journal of Computer Applications. find
  28. Sunil Regmi and Bal Krishna Bal. 2021. An end-to-end speech recognition for the Nepali language. In Proceedings of the 18th International Conference on Natural Language Processing (ICON), pages 180–185, National Institute of Technology find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sharma-2026-nwacha-muna-devanagari-speech-corpus/