400ms Voice AI: What It Takes

Developer Nick Tikhonov shared a Show HN project (122 upvotes) detailing how he built a voice agent averaging ~400ms end-to-end latency — from phone stop to first syllable — with a complete STT → LLM → TTS pipeline, clean barge-ins, and no precomputed responses.

What Actually Moved the Needle

Semantic End-of-Turn Detection: VAD alone fails for natural conversation. You need semantic understanding of when someone is truly done speaking
Streaming is Non-Negotiable: Sequential pipelines are dead on arrival. STT → LLM → TTS must all stream
TTFT Dominates: Groq's ~80ms time-to-first-token was the single biggest performance win
Geography Over Prompts: Colocating all components mattered more than any prompt optimization

The Core Loop

The system reduces to two states — speaking vs. listening — and two critical transitions: cancel instantly on barge-in, respond instantly on end-of-turn. These transitions define the entire user experience. Voice is fundamentally a turn-taking problem, not a transcription problem.

Open Source

The project is available on GitHub as 'shuo'. For developers building real-time voice AI systems, this implementation offers a practical, battle-tested reference for achieving sub-500ms conversational latency.

#voice-agent

Show HN: Building a Sub-500ms Latency Voice Agent from Scratch