Multilingual Low-Latency Emergency VoIP System Using LLM for Speech Reconstruction and Blockchain for Secure Data Archiving

Emergency VoIP communications frequently suffer from packet loss, jitter, and background noise, which degrade speech quality, delay dispatcher response, and threaten the reliability of emergency services. This paper presents a blockchain-governed, low-latency communication response framework powered by a Large Language Model (LLM), leveraging GPT-4o-mini for semantic speech reconstruction and dynamic emergency prioritization. The proposed architecture integrates multilingual speech-to-text and translation via Google APIs, a Facebook AI Similarity Search (FAISS)-indexed retrieval system for context-driven LLM inference, and severity classification for dispatcher triaging. To ensure tamper-proof data governance, the system incorporates AES-256-CBC encryption, Shamir’s Secret Sharing, and decentralized transcript storage on Arweave, with access mediated by Solana-based multi-signature wallets and program-derived addresses. Quantitative evaluations confirm that the proposed framework meets low-latency standards with sub-1.5 second one-way latency—significantly outperforming prior VoIP-based multilingual emergency systems, which typically report 3–5 second delays. Unlike conventional pipelines, it leverages asynchronous blockchain storage and optimized LLM-driven reconstruction to achieve high responsiveness without sacrificing Conceptual Precision, BLEU, or ROUGE accuracy.