1 Assistant Professor, Department of Information Technology, V.S.B Engineering College, Karur. Email: anandancse@gmail.com
2 Assistant Professor, Department of Artificial Intelligence and Data Science, V.S.B Engineering College, Karur. Email: maalini.cse@gmail.com
3 UG Scholar, Department of Artificial Intelligence and Data Science, V.S.B Engineering College, Karur. Email: vignesh2005mohan@gmail.com
4 UG Scholar, Department of Artificial Intelligence and Data Science, V.S.B Engineering College, Karur. Email: aafahim003@gmail.com
5 UG Scholar, Department of Artificial Intelligence and Data Science, V.S.B Engineering College, Karur. Email: muhamedasfack.s2005@gmail.com
6 UG Scholar, Department of Artificial Intelligence and Data Science, V.S.B Engineering College, Karur. Email: thilag48@gmail.com
Received: 20th Feb, 2026 | Revised: 4th Mar, 2026 | Accepted: 25th Mar, 2026 | Available Online: 10th Apr, 2026
Large language models (LLMs) have shown great promise in applications of legal question answering and contract analysis; however, the empirical study of hallucination has shown that it takes place in 23-38% of cases in regard to the legal facts, laws, and citations that are produced, thus making them not applicable to legal purposes. In this research, we introduce JurisGraph—an innovative concept for addressing the issue of legal inference hallucination using the neuro-symbolic artificial intelligence framework, where the inference is grounded in a knowledge graph. JurisGraph is based on a three-layer approach that includes the inclusion of statutory rules, case-law connections, and jurisdictional constraints into a knowledge graph as the first layer, logic control of triples in the graph relative to model-generated outputs as the second layer, and language generation by the model in response to conditional evidence from the graph as the third layer. What is important about the JurisGraph is that it guarantees grounding of all legal assertions generated in a factual knowledge graph. The results on the LegalBench and CUAD benchmarks demonstrated 40% hallucination reduction relative to RAG, and 91.3% citation accuracy achieved. Factual accuracy is ensured by the use of the module for neuro-symbolic reasoning, which has a positive influence on factual fidelity according to the results of ablation analysis, even though it was rather subtle. Moreover, this research presents JurisGraph that uses LLMs to carry out some crucial activities in law by adopting a scientific approach that allows for factual accuracy.
Keywords: AI Voice Agent, Telephony Calls, Noise Reduction, Indian Telephony, Speech Processing, Large Language Models.
How to cite this article: Anandan D, Maalini D, Vignesh M, Mohamed Fahim AA, Muhamed Asfack S, Thilag V. Ai Voice Agent For Noisy Indian Telephony Calls. Int J Drug Deliv Technol. 2026;16(29s):59-68. DOI: 10.25258/ijddt.16.29s.8
Source of support: Nil.
Conflict of interest: The authors declare no conflict of interest.