How a developer cut 2 seconds of dead air from a voice AI agent's response time

A software developer diagnosed a critical latency problem in a voice AI system where callers experienced roughly two seconds of silence before the agent spoke, causing them to talk over the greeting. By adding timestamps at each pipeline stage, the developer found that speech-to-text, LLM processing, and text-to-speech together accounted for the full delay. Three key inefficiencies were identified: waiting for the LLM to finish its entire response before sending text to the speech synthesizer, opening a fresh TTS connection on every turn, and providing no immediate audio acknowledgement to the caller. The fixes involved streaming LLM output sentence-by-sentence into TTS, keeping the synthesizer connection warm in advance, and playing a short filler phrase while the full response was being generated. These changes eliminated the dead-air gap without requiring faster AI models or upgraded infrastructure.