31, October 2026

Domain-Specific Noise-Robust Hindi Automatic Speech Recognition Using Fine-Tuned IndicWav2Vec

VOLUME-12; ISSUE-10; OCT-2026

Author(s): 1. David Raju Kolluri, 2. Kotti Bhuvana Teja

Authors Affiliations:

¹ Associate Professor, HITAM, Hyderabad, India

² UG Scholar, Department of Computer Science and Engineering (Data Science), Hyderabad Institute of Technology and Management, Hyderabad, India

DOIs:10.2015/IJIRMF/202610027     |     Paper ID: IJIRMF202610027


Abstract
Keywords
Cite this Article/Paper as
References

Automatic speech recognition (ASR) systems transform spoken utterances into written text, supporting applications across sectors such as agriculture, financial services, and digital customer assistance. Recognition accuracy may decline when speech contains environmental noise, variations in pronunciation, differences between speakers, or specialized vocabulary. This research examines domain-oriented adaptation of the pretrained IndicWav2Vec model for Hindi speech using noisy recordings from the RESPIN dataset. Independent models were adapted for the Agriculture and Finance domains, with 20,000 utterances allocated for training and 2,500 for validation in each domain; a further 2,500 utterances were used for testing. Performance was assessed through Word Error Rate (WER) and Character Error Rate (CER) on in-domain and cross-domain speech. In the Finance experiments, the adapted system recorded WER and CER values of 25.34% and 11.02%, respectively, while the baseline produced 30.08% and 13.14%. Agriculture results showed a different pattern: CER improved from 13.26% to 12.64%, but WER increased slightly from 29.90% to 30.32%. These findings indicate that adapting Hindi ASR to a specific domain can improve recognition, although the magnitude and consistency of the gains vary across domains.

Hindi Automatic Speech Recognition, IndicWav2Vec, Domain Adaptation, RESPIN Dataset, Noisy Speech, Word Error Rate, Character Error Rate.

David Raju Kolluri, Kotti Bhuvana Teja (2026); Domain-Specific Noise-Robust Hindi Automatic Speech Recognition Using Fine-Tuned IndicWav2Vec, International Journal for Innovative Research in Multidisciplinary Field, ISSN(O): 2455-0620, Vol-12, Issue-10, Available on –   https://www.ijirmf.com/

 

  1. Baevski, A., Zhou, Y., Mohamed, A., and Auli, M. (2020). wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in Neural Information Processing Systems, 33, 12449–12460. https://arxiv.org/abs/2006.11477
  2. Javed, T., Doddapaneni, S., Raman, A., Bhogale, K. S., Ramesh, G., Kunchukuttan, A., Kumar, P., and Khapra, M. M. (2022). Towards building ASR systems for the next billion users. Proceedings of the AAAI Conference on Artificial Intelligence, 36(10), 10813–10821. https://doi.org/10.1609/aaai.v36i10.21327
  3. Kumar, S., et al. (2025). RESPIN-S1.0: A read speech corpus of 10000+ hours in dialects of nine Indian languages. Advances in Neural Information Processing Systems, 38. https://doi.org/10.52202/085713-4928
  4. Graves, A., Fernández, S., Gomez, F. J., and Schmidhuber, J. (2006). Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd International Conference on Machine Learning (pp. 369–376). https://doi.org/10.1145/1143844.1143891
  5. AI4Bharat. (n.d.). IndicWav2Vec. https://indicnlp.ai4bharat.org/indicwav2vec/

 

 

Views: 21

Download Full Paper

Download PDF No. of Downloads:0 | No. of Views: 18