Nwāchā Munā (/nwa:tSa: muna:/) means “speech corpus” or "collection of speech" in Nepal Bhasha (Newari).

Authors: Rishikesh Kumar Sharma, Safal Narshing Shrestha, Jenny Poudel, Rupak Tiwari, Arju Shrestha, Rupak Raj Ghimire, Bal Krishna Bal
Information and Language Processing Research Lab (ILPRL)
Kathmandu University, Nepal

Nepal Bhasha (Newari), an endangered language of the Kathmandu Valley, remains digitally marginalized due to the severe scarcity of annotated speech resources. In this work, we introduce Nwāchā Munā, a newly curated 5.39-hour manually transcribed Devanagari speech corpus for Nepal Bhasha, and establish the first benchmark using script-preserving acoustic modeling.

We investigate whether proximal cross-lingual transfer from a geographically and linguistically adjacent language (Nepali) can rival large-scale multilingual pretraining in an ultra-low-resource Automatic Speech Recognition (ASR) setting. Fine-tuning a Nepali Conformer model reduces the Character Error Rate (CER) from a 52.54% zero-shot baseline to 17.59% with data augmentation, effectively matching the performance of the multilingual Whisper-Small model despite utilizing significantly fewer parameters.

Our findings demonstrate that proximal transfer within South Asian language clusters serves as a computationally efficient alternative to massive multilingual models. We openly release the dataset and benchmarks to digitally enable the Newari community and foster further research in Nepal Bhasha.

Preprint: arXiv: Read

Full Paper: To be linked upon publication at CHiPSAL @ LREC 2026

Dataset and Models: Nwāchā Munā — Hugging Face Collection

Nepal Bhasha (Newari) is spoken by over 860,000 people in Nepal and holds official working language status in Bagmati Province. Despite a six-century-old history, it is classified by UNESCO as a "definitely endangered" language — a stark contrast to its cultural significance. In a country with 124 official languages, the absence of digital speech tools for indigenous languages like Newari widens an already deep divide.

This work takes a direct step toward closing that gap. We curate the first large-scale Devanagari-script speech corpus for Nepal Bhasha and establish a rigorous ASR benchmark. Crucially, we ask a question with broad implications for the region: can a model trained on a neighboring language like Nepali — sharing the same script and acoustic neighborhood — substitute for the massive compute and data demands of multilingual models like Whisper?

A Conformer model pre-trained on Nepali and fine-tuned on just 5.39 hours of Newari speech achieves a CER of 18.72%, matching the fine-tuned Whisper-Small model (18.76%) — a model 5× larger and backed by orders of magnitude more multilingual training data. Shared Devanagari script and acoustic similarity between the two languages make this possible.

Augmenting the 5.39-hour corpus (speed perturbation, pitch shifting, noise injection) expanded the effective training data to 23.05 hours and pushed the best CER down to 17.59%, the top result across all experiments. When data is scarce, augmentation can close the gap that raw scale cannot.

Despite sharing the Devanagari script, the zero-shot NepConformer achieves only 52.54% CER on Newari — confirming that phonological distinctiveness between the two languages requires explicit fine-tuning. Script sharing is a useful head-start, not a free pass.

Freezing the Nepali encoder and fine-tuning only the decoder yields 18.77% CER — nearly identical to full model fine-tuning. This suggests the Nepali acoustic encoder already captures representations transferable to Newari, reducing the compute cost of adaptation without sacrificing quality.

A semi-supervised experiment incorporating 9.33 hours of filtered broadcast audio (radio, podcasts, TV) degraded performance to 19.83% CER. In ultra-low-resource settings, domain alignment matters more than raw data volume. Noisy, out-of-domain pseudo-labels can outweigh the gains of additional data.

Integrating a 5-gram KenLM via shallow fusion reduced WER by ~11.7% relative, reflecting better word-level consistency. However, CER slightly increased, as the language model occasionally overrides phonetically valid but non-standard local Newari forms — a known tension in agglutinative, low-resource languages.

Data was collected from 18 native speakers across Banepa, Dhulikhel, Panauti, and Patan, stratified by age and gender to ensure community representation.

If you use Nwāchā Munā in your research, please cite:

@inproceedings{sharma2026nwachamuna,
  title     = {Nw\={a}ch\={a} Mun\={a}: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR},
  author    = {Sharma, Rishikesh Kumar and Shrestha, Safal Narshing and Poudel, Jenny
               and Tiwari, Rupak and Shrestha, Arju and Ghimire, Rupak Raj and Bal, Bal Krishna},
  booktitle = {Proceedings of CHiPSAL @ LREC2026},
  year      = {2026},
  url       = "https://arxiv.org/abs/2603.07554",
}