Abstract
Emergency communication systems often suffer from disruptions due to packet loss, bandwidth limitations, and poor VoIP signal quality, making it difficult for dispatchers to assess the urgency of a situation. Additionally, victims in distress may struggle to convey critical information due to panic, speech disorders, or background noise, leading to incomplete or ambiguous communication. This paper proposes an LLM-based real-time speech reconstruction and emergency call prioritization system to address these challenges. The proposed system integrates real-time audio transcription with retrieval-augmented input conditioning using TF-IDF and FAISS indexing to retrieve relevant emergency transcripts, enhancing contextual understanding. Unlike traditional Retrieval-Augmented Generation (RAG) models, this approach eliminates the need for fine-tuning, making it lightweight and efficient for real-time applications. The reconstructed speech output is further classified using a GPT-based severity detection model, ensuring that critical emergencies receive immediate attention via Twilio and AssemblyAI APIs. Experimental evaluations demonstrate high Conceptual Precision and favorable BLEU and ROUGE scores, indicating strong alignment with real-world emergency scenarios. The results highlight the system’s potential to improve emergency response efficiency, optimize call triaging, and reduce communication delays in high-risk situations