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

Rishikesh Kumar Sharma; Safal Narshing Shrestha; Jenny Poudel; Rupak Tiwari; Arju Shrestha; Rupak Raj Ghimire; Bal Krishna Bal

cs.CLMar 8, 2026

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

Authors: Rishikesh Kumar Sharma, Safal Narshing Shrestha, Jenny Poudel, Rupak Tiwari, Arju Shrestha, Rupak Raj Ghimire, Bal Krishna Bal

Organizations: Information and Language Processing Research Lab, Kathmandu University

Abstract

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 from Nepali language 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.

Figures & tables

Condition	Sentence Structure
Before	\nep ”छगु ब्वसय् २१% ब्व दु।” chagu bwasay 21% bwa du The sentence contains 21 as a number.
After	\nep छगु ब्वसय् नीछगू सयेक ब्व दु chagu bwasay nīchagū sayeka bwa du The sentence contains 21 in textual form

Table 1: Example of Text Normalization for Nepal Bhasha ASR

Metric	Count
Total Sentences	5,727
Total Words	27,644
Unique Words in Corpus	8,599

Table 2: Nwāchā Munā Corpus Overview

Sentence Length	Count
Short (0–5 words)	4,323
Medium (6–10 words)	1,073
Long ( $>10$ words)	331

Table 3: Sentence Length Distribution in Nwāchā Munā Corpus

Metric	Value
Total Utterances	5,727
Total Duration	5.39 hours
Mean Utterance Length	3.39 sec

Table 4: Overview of the Nwāchā Munā

Age Group	Number of Speakers
16–25	6
26–30	2
30+	10
Total Speakers	18

Table 5: Age Distribution of Speakers in the Nwāchā Munā Corpus

Gender	Number of Speakers
Male	10
Female	8
Total Speakers	18

Table 6: Gender Distribution of Speakers in the Nwāchā Munā Corpus

Split	Hours	Utterances
Train	4.31	4575
Validation	0.54	576
Test	0.54	576

Table 7: Dataset split for Newari speech corpus

Strategies (model details)	Dataset	CER (%)	WER (%)
Zero-Shot NepConformer	-	52.54	98.68
Semi-supervised NepConformer	5.39 hours labelled + 9.33 hours Pseudo labeled	19.83	58.28
NepConformer + 5-gram KenLM (Shallow fusion, $\lambda=0.4$ , $\beta=1.5$ , beam 128)	5.39 hours labelled	19.75	48.46
Decoder-only NepConformer (Encoder frozen)	5.39 hours labelled	18.77	53.47
Whisper-Small (Fine-tuned on base data)	5.39 hours labelled	18.76	51.44
NepConformer (Fine-tuned on base data)	5.39 hours labelled	18.72	54.34

Table 8: Summary of Results for Evaluated Modeling Strategies

Features	Nepali	Newari
Script	Devanagari	Devanagari
Nasalization Markers	Yes	Yes (more enhanced)
Agglutination	Moderate	High
Halant Usage	Standard	High Morphological Load

Table 9: Orthographic and Phonological Comparison of Nepali and Newari

Model	Parameter (Million)	VRAM Used (GB)
NepConformer	30.54	8.56
Whisper(small)	244	12.38

Table 10: Computational Resource and Parameter Comparison

Error Type	% of total errors
Word Boundary Errors	35%
Halant/Cluster Errors	25%
Nasalization & Anusvara Confusion	20%
Lexical Substitution	15%
Function Word & Particle Error	5%

Table 11: Error Frequency Table

No.	Reference	Prediction	Error Analysis
(i)	\nep तस्सकं न्ह्याइपुसेच्वं सीम बत्तिं चैं छन्दत म्याहला च्वंगु tassakaṃ nhyāipusecvaṃ sīma battiṃ caiṃ chandata myāhalā cvaṃgu It was a lot of fun sitting even in the dim light, singing your song	\nep तस्स कां न्येयेपुसं सिमा पत्ति चैं छंत मां ला च्वं tassa kāṃ nyeyepusaṃ simāpatta caiṃ chaṃta māṃ lā cvaṃ	Phonetic substitutions and word distortions; nasal markers and vowels altered within agglutinated structures.
(ii)	\nep दइ धका नं सिल हे daï dhakā naṃ sila he	\nep दइ ध काः नं सिल हे daï dha kāḥ naṃ sila he	Insertion of \nep (ः) (Visarga), reflecting minor phonetic misalignment.
	They say it is known that it will fall
(iii)	\nep फोहोर जुलकि लाकां सिले माः phohora julaki lākāṃ sile māḥ If the shoe is dirty, it must be washed	\nep पर जुलकि लाका सि ने माः para julaki lākā si ne māḥ	Nasal suffix deletion and lexical substitution; final \nep (ं) dropped, with minor token distortion.
(iv)	\nep आः थ्व सु	\nep आाख्वसु	Word boundary deletion; independent tokens merged due to agglutinative structure.
	āḥ thva su	āākhvasu

Table 12: Qualitative Error Analysis of Nwāchā Munā (underline in prediction column represents the error segment)

Explore similar work

Neural Speaker Diarization via Multilingual Training: Evaluation on Low-Resource Nepali-Hindi Speech
Jun 21, 2026Samip Neupane, Sandesh Pokhrel, Sandesh Pyakurel +1Low-Resource Language Processing Speaker Diarization
Department of Electronics and Computer Engineering, Pulchowk Campus, Institute of Engineering, Lalitpur, Nepal
Sometin Beta Pass Notin (SBPN): Improving Multilingual ASR for Nigerian Languages via Knowledge Distillation
May 18, 2026Sewade OgunLow-Resource Language Processing Low-Resource Speech Recognition
Benchmarking Linguistic Adaptation in Comparable-Sized LLMs: A Study of Llama-3.1-8B, Mistral-7B-v0.1, and Qwen3-8B on Romanized Nepali
Mar 25, 2026Ananda Rimal, Adarsha RimalLow-Resource Language Processing Domain Adaptation for LLMs
Nepal Engineering College · Dept. of Computer Science & Engineering, Nepal Engineering College · Tribhuvan University +1

Jun 21, 2026J/K move · Enter open · S save

Neural Speaker Diarization via Multilingual Training: Evaluation on Low-Resource Nepali-Hindi Speech

Samip Neupane, Sandesh Pokhrel, Sandesh Pyakurel, Basanta Joshi

Department of Electronics and Computer Engineering, Pulchowk Campus, Institute of Engineering, Lalitpur, Nepal

Low-Resource Language Processing Speaker Diarization Cross-Lingual Transfer

Open paper AskPDF arXiv

Speaker diarization, the task of determining "who spoke when" in a multi-speaker recording, is a critical component in applications such as meeting transcription, accessibility tools, and multilingual information retrieval. While end-to-end neural diarization systems have achieved strong performance for English and other high-resource languages, their effectiveness degrades substantially for underrepresented languages where annotated speech data is scarce. This paper investigates speaker diarization for low-resource Nepali-Hindi speech through a multilingual training approach, comparing two modern architectures: EEND with encoder-decoder attractors (EEND-EDA) and EEND with Perceiver-based attractors (DiaPer). Both models are trained on a multilingual corpus combining English speech from LibriSpeech, diverse speaker recordings from VoxCeleb, and separately collected Nepali and Hindi audio, a setup designed to reduce language bias and encourage cross-lingual generalization. We evaluate both models across 2-speaker, 3-speaker, 4-speaker, and mixed-speaker scenarios on LibriSpeech, VoxCeleb, and Nepali-Hindi (NeHi) test sets. DiaPer achieves stronger overall performance than EEND-EDA, particularly in more challenging multi-speaker conditions, obtaining DERs of 3.28%, 2.02%, 4.05%, and 4.76% on NeHi 2-speaker, 3-speaker, 4-speaker, and mixed-speaker settings, respectively, compared to 1.50%, 9.68%, 16.17%, and 11.19% for EEND-EDA. These results demonstrate the viability of Perceiver-based end-to-end neural diarization for low-resource multilingual speech processing.

Abstract

Explore similar work

Neural Speaker Diarization via Multilingual Training: Evaluation on Low-Resource Nepali-Hindi Speech

Sometin Beta Pass Notin (SBPN): Improving Multilingual ASR for Nigerian Languages via Knowledge Distillation

Benchmarking Linguistic Adaptation in Comparable-Sized LLMs: A Study of Llama-3.1-8B, Mistral-7B-v0.1, and Qwen3-8B on Romanized Nepali