RELAY — Real-Time Voice AI Support

RELAY is a real-time voice AI support system that prevents customers from repeating themselves during human escalation. THE PROBLEM Customers may interrupt to correct details: “I tried logging in TODAY, not yesterday.” AI can miss the correction and retain outdated information. During escalation, agents may receive a transcript without the latest intent, prior actions, or transfer reason. Customers repeat their story, increasing frustration and wasting time. HOW IT WORKS • Real-time voice capture processes speech for natural language understanding. • Barge-in detection stops AI speech when customers interrupt (94% stated accuracy). • Context preservation recognizes corrections, updates session state, and reasons from current details. • Intelligent escalation detects requests for human help. • Structured handoff shares intent, corrections, prior actions, context, and escalation reason. KEY METRICS 87% less customer repetition; 42% faster escalation resolution; 28% improved agent efficiency; 98% session state consistency; 180ms transcription latency; and 20+ turns supported without reported context loss. These are stated project metrics and require controlled testing for validation. TECHNICAL STACK FastAPI and WebSockets enable full-duplex voice interaction. AssemblyAI Voice Agent API supports speech processing, Python manages session state and semantic context, and the Web Audio API captures microphone input. WHY IT MATTERS Lost context makes customers repeat information and agents reconstruct previous events. RELAY aims to make escalation a continuation of the same conversation. DIFFERENTIATION Unlike generic chatbots focused on individual answers, RELAY is designed for support workflows. It treats corrections as updates, preserves context across turns, and coordinates human escalation.