Definition

Voice AI Pipeline

A voice agent chains three models: speech-to-text transcribes the caller, an LLM reasons and drafts a reply, and text-to-speech renders it as audio. Each stage adds latency, and human conversation expects a reply within a few hundred milliseconds, so every stage must stream into the next. Around 700 milliseconds to first audio feels like a natural pause; below 500 feels real-time. Silence during tool calls is deadly.