Nwāchā Munā: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR
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 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.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The main contribution of this paper is the introduction of Nwāchā Munā, a Devanagari speech corpus for Nepal Bhasha, and the demonstration that proximal transfer from Nepali can effectively enhance ASR performance in ultra-low-resource settings. This research not only addresses a critical gap in the availability of resources for endangered languages but also showcases innovative methodologies that could influence future work in low-resource ASR systems.
Methodology
The methodology is robust, comprising a well-structured data collection process that includes both textual and acoustic resources. The authors employ a dual-modal strategy for audio acquisition and utilize a variety of data augmentation techniques to enhance model robustness. The comparative experimental framework effectively evaluates the performance of different transfer learning strategies, providing insights into the effectiveness of proximal transfer versus large multilingual models. The use of a semi-supervised learning approach to leverage unlabeled data is particularly commendable, although it appears to have mixed results.
Results & evaluation
The experimental setup is thorough, with a clear focus on character error rates (CER) as the primary metric. The results indicate significant improvements through fine-tuning and data augmentation, showcasing the potential of the proposed methods. The authors provide a detailed analysis of the experimental results, including a comprehensive error analysis that highlights the challenges posed by Newari's morphological structure. However, the reliance on CER alone may limit the understanding of model performance in practical applications.
Reproducibility
The paper provides sufficient details regarding the model architecture, training strategies, and data collection processes, which aids reproducibility. However, the absence of a publicly available code repository or demo URL limits the ease of reproducing the results. The authors mention the use of specific tools and frameworks, but without access to the actual implementation, full reproducibility may be challenging.
Limitations
The primary limitation is the small size of the Nwāchā Munā dataset, which may not fully capture the acoustic variability found in spontaneous speech. Additionally, the focus on read speech and formal sentences may not represent the complexities of everyday language use. The external language model's limited coverage of conversational forms also poses challenges for practical deployment.
Broader impact
This work has significant implications for the preservation and digital empowerment of endangered languages, particularly Nepal Bhasha. By providing a high-quality speech corpus and establishing benchmarks, the research fosters further development in ASR technologies for underrepresented languages. The findings could serve as a blueprint for similar initiatives in other linguistic communities, promoting inclusivity in the technological landscape. The main contribution of this paper is the introduction of Nwāchā Munā, a Devanagari speech corpus for Nepal Bhasha, and the demonstration that proximal transfer from Nepali can effectively enhance ASR performance in ultra-low-resource settings. This research not only addresses a critical gap in the availability of resources for endangered languages but also showcases innovative methodologies that could influence future work in low-resource ASR systems.
Author of this paper? Send a correction.