Robust and Scalable Speech Recognition Framework for Low-Resource Languages Using Adaptive Deep Transfer Learning and Noise-Aware Multilingual Modeling

Tadi Chandrasekhar, S. Srinivasulu Raju, Rajesh Kumar K., R. V. Kavya, D. B. K. Kamesh, Deepa Malini M.

2025

Abstract

Low-Resource languages continue to experience difficulties in building accurate automatic speech recognition (ASR) systems because of a lack of data, phonetic diversity and lack of representation in worldwide linguistic resources. In this paper, we present a novel and scalable speech recognition architecture that uses a combination of adaptive deep transfer learning and multilingual model to address these problems. It exploits self-supervised learning techniques, pre-trained cross-lingual embeddings and dynamic adapter modules to transfer knowledge across distant language families in a low-resourced setting with a low reliance on large labeled corpora. To increase deployability, the framework integrates model compression techniques, such as pruning and quantization, for real-time inferencing on the edge devices. Furthermore, noise-aware training and sharp feature preserving methods are used to enhance the performance in both noisy and tonal language scenes. Experiments on newly collected datasets from underrepresented languages confirm the superiority of the model regarding both accuracy and generalisability, as well as computational cost, in contrast to existing methods. The proposed method provides a new benchmark in terms of inclusive, state-of-the-art speech recognition systems that can be configured for multiple linguistic and operational settings.

Paper Citation

in Harvard Style

Chandrasekhar T., Raju S., K. R., Kavya R., Kamesh D. and M. D. (2025). Robust and Scalable Speech Recognition Framework for Low-Resource Languages Using Adaptive Deep Transfer Learning and Noise-Aware Multilingual Modeling. In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies - Volume 1: ICRDICCT`25; ISBN 978-989-758-777-1, SciTePress, pages 632-639. DOI: 10.5220/0013870500004919

in Bibtex Style

@conference{icrdicct`2525,
author={Tadi Chandrasekhar and S. Raju and Rajesh K. and R. Kavya and D. Kamesh and Deepa M.},
title={Robust and Scalable Speech Recognition Framework for Low-Resource Languages Using Adaptive Deep Transfer Learning and Noise-Aware Multilingual Modeling},
booktitle={Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies - Volume 1: ICRDICCT`25},
year={2025},
pages={632-639},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013870500004919},
isbn={978-989-758-777-1},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies - Volume 1: ICRDICCT`25
TI - Robust and Scalable Speech Recognition Framework for Low-Resource Languages Using Adaptive Deep Transfer Learning and Noise-Aware Multilingual Modeling
SN - 978-989-758-777-1
AU - Chandrasekhar T.
AU - Raju S.
AU - K. R.
AU - Kavya R.
AU - Kamesh D.
AU - M. D.
PY - 2025
SP - 632
EP - 639
DO - 10.5220/0013870500004919
PB - SciTePress