PP-MET: A REAL-WORLD PERSONALIZED PROMPT BASED MEETING TRANSCRIPTION SYSTEM

Abstract

Accurate capture and structured analysis of meeting discussions remain critical challenges in knowledge-driven organizations, particularly where confidentiality and contextual precision are required. This paper presents PPMET (Personalized Prompt-Based Meeting Transcription System), a fully on-premises, AI-integrated framework designed to automate the complete post-meeting intelligence pipeline. The system transforms raw audio recordings into speaker-attributed transcripts through a multi-stage architecture combining speech activity detection, speaker diarization, and Transformer-based automatic speech recognition. For multilingual scenarios, an intermediate neural translation module standardizes outputs into a unified language to support consistent downstream processing. The resulting transcript is segmented into semantically coherent chunks, converted into dense vector embeddings, and indexed within a persistent vector database to enable efficient semantic retrieval. PP-MET further employs personalized prompt engineering to enhance contextual relevance during summarization and implements a Retrieval-Augmented Generation mechanism that grounds responses in retrieved transcript segments, thereby minimizing generative inconsistencies. The architecture follows a decoupled, multithreaded design to maintain interface responsiveness while executing computationally intensive tasks. Experimental evaluation on consumer-grade hardware demonstrates reliable transcription fidelity, structured summary generation, and context-aware question answering. The proposed framework establishes a secure, scalable solution for transforming unstructured spoken discourse into searchable, actionable institutional knowledge.