Streaming TTS & the Latency Budget

The metric that defines streaming TTS isn't audio fidelity — it's TTFB: time to first byte (or first audible sample). A system that synthesizes the full utterance then plays is dead for agents: the user waits through the entire generation while silence says 'broken'.

The mental model: conversational response time needs to land under ~1 second; TTS TTFB is the tail of that budget, after ASR and the LLM. Perfect audio delivered at 1.8s loses to good audio delivered at 250ms — humans forgive texture, they don't forgive dead air.