Problem Statement: Misinformation spreads rapidly during live events (e.g., debates, streams), often outpacing post-hoc verification. While Speech-to-Text (STT) exists, the challenge lies in processing infinite audio streams. A system must intelligently "chunk" continuous audio into semantic sentences (Audio Segmentation), transcribe them with low latency, and verify claims against a knowledge base without breaking the live flow.
Project Goal: The goal is to develop a "Near Real-Time" Analysis Pipeline. The system will ingest a live audio stream (Mic or WebSocket), segment it using Voice Activity Detection (VAD), transcribe speech to text, and trigger an automated fact-checking agent (LLM or FactCheck API) to display a "Truth Overlay" on a dashboard.
Core Tasks:
- Audio Stream Processing: Implement a buffering system using WebSockets that handles continuous audio data. Crucially, integrate Voice Activity Detection (VAD) to slice the stream into grammatically valid "chunks" for processing.
- Transcription Pipeline: Integrate a high-speed STT engine to convert audio chunks to text on the fly.
- Claim Verification Agent: Develop a lightweight NLP module that detects "check-worthy" claims and verifies them using an external API (e.g., Google FactCheck Claims API or an LLM with Search access).
- Live Dashboard: Build a React frontend that displays the rolling transcript and highlights "Verified" vs. "Disputed" statements with minimal delay.
Research Focus:
The research focus lies on a System Performance Evaluation. The student must benchmark the End-to-End Latency (Time-to-Verification) and analyze the trade-off between Segmentation Length (waiting for a full sentence) vs. Transcription Accuracy (Word Error Rate).
Technologies: Python (FastAPI/WebSocket); Voice Activity Detection (Silero VAD); Streaming STT (Whisper / Google); LLMs (LangChain); React
Tags: Multimedia Stream Processing; Audio Analysis; Natural Language Processing; Speech Recognition; Real-Time Systems; Kalchgruber
