BUILDING A REAL-TIME AI SYSTEMFOR RECOGNIZINGANDSUMMARIZING IT CONVERSATIONS USING WHISPER ANDLLAMA3

Abstract

As remote and hybrid work environments become the industry standard, thevolume of virtual IT communications has surged. Consequently, manuallycapturing,transcribing, and summarizing highly technical discussions has proventobeinefficient, error-prone, and resource-intensive. This project addressesthesechallenges by introducing an end-to-end, automated Artificial Intelligencepipelinedesigned to capture real-time audio and extract structured meeting intelligencewithminimal human intervention. The proposed system architecture is divided into three core processingphases.First, it employs Faster-Whisper coupled with optimized Voice ActivityDetection(VAD) for low-latency, highly accurate Speech-to-Text (STT) transcription, whichisspecifically configured to recognize complex IT terminologies. Second,Pyannote.audio is integrated to perform precise speaker diarization, clusteringaudioframes to distinguish individual participants. Third, the systemleverages acustomfine-tuned Large Language Model (LLM)—specifically Llama-3-8B-Instructoptimized via Low-Rank Adaptation (LoRA)—to semantically analyze thediarizedtext. This LLM engine is trained to filter out conversational noise and automaticallygenerate a structured post-meeting report encompassing participant identification, keydiscussion points, technical vocabulary, and actionable items. To support seamless real-time operations, the backend is built on FastAPI witharobust Dual-WebSocket architecture. This design effectively bypasses strict networkconstraints (e.g., CORS and tunneling limitations) and facilitates uninterrupted,asynchronous bidirectional data streaming. On the frontend, an enterprise-gradeSplit-Panel Dashboard is developed. It uniquely incorporates Document Picture-in-Picture(PiP) technology, ensuring that live captions remain persistently visible (always-ontop) while users seamlessly switch tabs or share screens during presentations.Ultimately, by combining state-of-the-art NLP models with a fault-tolerant webinfrastructure, this project delivers a highly scalable and practical solutionthatsignificantly modernizes enterprise communication workflows.