Loading article...
Loading article...
Stream/Area:
-
Subject:
-
Paper ID:
IJSREM68875
Volume:
10 |
11 |
Nov 2026
Pages:
1-4
Publication Frequency:
Monthly (12 issues per year)
License:International Journal of Scientific Research in Engineering and Management follows the Creative Commons Attribution 4.0 (CC BY 4.0) License, allowing free use and distribution with proper attribution.

Artificial Intelligence and Machine Learning
Vivekananda College of Engineering and Technology

Artificial Intelligence and Machine Learning
Vivekananda College of Engineering and Technology

Artificial Intelligence and Machine Learning
Vivekananda College of Engineering and Technology

Artificial Intelligence and Machine Learning
Vivekananda College of Engineering and Technology

Artificial Intelligence and Machine Learning
Vivekananda College of Engineering and Technology
Tulu is a Dravidian language of coastal Karnataka and northern Kerala. It is normally written in Kannada script, and very little of it exists as digital text. Because of this, mainstream AI models tend to reply to Tulu questions in Kannada and to invent facts they have never seen. This paper presents VoxIntel: A Multimodal Intelligent Assistant For Tulu Language, supporting conversation, translation, lessons and speech on a standard laptop without any cloud AI service or API key. The system combines a Retrieval-Augmented Generation (RAG) knowledge base of 2,069 native-reviewed records with a locally hosted Gemma 3 4B model fine-tuned with QLoRA on a 12,029-entry Tulu dataset. When a user enters a query, the RAG knowledge base is searched first using embedding similarity. If a matching record is found, its reviewed answer is returned; only when no record matches does the Gemma 3 4B model generate the response. A Kannada detector, an automatic rewrite step and an output guard keep generated replies in Tulu. On 119 unseen replies the fine-tuned model scored 45.7 chrF against 20.2 for the earlier model, and its word-level F1 rose from 0.154 to 0.482. Repeated-word loops disappeared. The router chose the correct answer source for 88 of 100 test messages, and 1,602 automated tests pass.
KEYWORDS
VoxIntel demonstrates the use of artificial intelligence to build an offline conversational assistant for the Tulu language. By checking a native-reviewed RAG knowledge base of 2,069 records first and using a Gemma 3 4B model fine-tuned on a 12,029-entry Tulu dataset only when no record matches, the system provides improved Tulu responses through chat, translation, learning and voice interaction. The new model scored 45.7 chrF against 20.2 for the earlier version, raised word-level F1 from 0.154 to 0.482 and removed repeated-word loops, while the router selected the correct source for 88 of 100 messages. Local processing also keeps conversations on the user’s device, preserving privacy. Overall, VoxIntel provides a practical approach to supporting Tulu communication and promoting the digital preservation of regional languages.
1. S. Prabhu, D. V, S. Kamath, M. H, K. K. Prakasha, and D. A. Dinesh, “Spoken Language Identification System for Detecting Coastal Karnataka Languages Using Transfer Learning and Multi-View Learning,” Engineering Research Express, vol. 8, no. 1, p. 015203, 2026. 2. S. Sani, S. V. Gangashetty, K. Samudravijaya, and A. K. Dubey, “Emotion-Semantic-Aware Neural Machine Translation Between Indo-Aryan and Dravidian Languages via Transfer Learning,” IEEE Access, vol. 14, pp. 4953–4969, 2026. 3. N. Telrandhe and R. Kadu, “Machine Translation Models for Low-Resource Indo Aryan Languages: A Comprehensive Review of Techniques, Challenges and Future Scope,” in Proc. 3rd DMIHER Int. Conf. Artificial Intelligence in Healthcare, Education and Industry (IDICAIHEI), Wardha, India, 2025, pp. 1–6. 4. A. K, A. C, and M. Balipa, “Natural Language Processing Model for Hindi to Tulu Translation,” in Proc. Int. Conf. Advances in Modern Age Technologies for Health and Engineering Science (AMATHE), Shivamogga, India, 2025, pp. 1–6. 5. M. Balipa, A. K, Ananya, and M. Murugappan, “A Rule-Based Machine Translation Framework for Low-Resource Language Pairs,” in Proc. 4th Int. Conf. Sentiment Analysis and Deep Learning (ICSADL), Bhimdatta, Nepal, 2025, pp. 969–974. 6. S. Shetty, S. Shetty, and N. V. Bangera, “TuluAgri Assist: Personalized Voice Assistant for Agriculture and Farming Support,” in Proc. IEEE Int. Conf. Distributed Computing, VLSI, Electrical Circuits and Robotics (DISCOVER), Mangalore, India, 2025, pp. 554–557. 7. M. Lau, “Alexa, Google Assistant, and Siri’s language options: how voice assistants reproduce monoglossic language ideologies,” AI & Society, vol. 40, pp. 6831–6846, 2025. 8. A. Mannonov, L. H. Jasim, A. S. Anvarovna, W. Suryasa, andA. Nayak, “Bridging Speech and Text using Multimodal Artificial Intelligence for Next-Gen Language Understanding,” in Proc. Int. Conf. Computational Innovations and Engineering Sustainability (ICCIES), Coimbatore, India, 2025, pp. 1–6. 9. M. N. C, R. Natarajan, and S. R, “Multimodal Hate Speech Detection in Dravidian Languages Using Text and Audio Features,” in Proc. Int. Conf. Communication and Signal Processing (ICCSP), Melmaruvathur, India, 2025, pp. 148–152. 10. S. Sidramappa, M. V. Reddy, and S. K, “Machine Learning-Driven System for Kannada-to-English Text Translation,” in Proc. 2nd Int. Conf. Artificial Intelligence and Knowledge Discovery in Concurrent Engineering (ICECONF), Chennai, India, 2025, pp. 1–6.
Publication Date
11/10/2026
Peer Review Type
Double-Blind Peer Review
Review Rounds: 2
Documents
Article Views
5
Total views since publication
PDF Downloads
4
Total downloads since publication
Geographic Reach
5
Digital Preservation: This article is permanently archived on Zenodo and complies with all data preservation standards across major indexing platforms.
Plagiarism Screening
This manuscript has been checked for plagiarism using TurnitinPlagiarismChecker software.
Conflict of Interest
The authors declare no conflict of interest related to this research work.
Funding Information
No funding information available for this research.
Related research in the same subject area
Loading similar articles...