What we learned shipping voice agents

Voice is unforgiving in a way chat is not. A 400ms pause in a chat window is invisible; the same pause on a call reads as the other person not listening.

Time to first token is the number people optimise, but the caller experiences transport, transcription, inference and synthesis. Optimising one stage while another regresses produces a system that benchmarks well and sounds worse.

Nobody listens to call recordings at volume. They read transcripts, search them, and quote them into tickets. Getting speaker attribution and timing right does more for the operations team than another point of audio quality.