voice-agents · davila7 — Claude Code skill: build low-latency voice agents

What voice-agents does

The skill lays out two ways to build a voice agent. Speech-to-speech on the OpenAI Realtime API sends audio straight to audio: lowest latency, emotion preserved, less control. The pipeline route runs STT→LLM→TTS, so every step stays inspectable and easier to debug, at the cost of added milliseconds.

Its core point: latency is the constraint. Each component contributes its own delay, and the total decides whether a conversation feels natural. So the skill tells you to measure and budget latency per component first, before picking models.

Inside are patterns for voice activity detection, barge-in detection and turn-taking, plus a list of anti-patterns. The sharp edges get their own section: turn detection that only listens for silence, long responses, background noise, STT errors. It pairs with agent-tool-builder, multi-agent-orchestration, llm-architect and backend.