The quarter-second that kills a voice agent
Latency budgets for AI voice, and why your SBC placement matters more than your model choice.
Lead image — The quarter-second that kills a voice agentREF: INS-ART-05JPG, PNG or WebP · 16:9 · min 1920×1080pxImage placeholder — supply asset with this REF
Editorial note: this piece is still an outline — the section list below is what it will cover. The full article is being drafted.
A voice agent that takes a quarter of a second too long to start speaking does not sound slow. It sounds unsure. Callers interrupt it, talk over it, and conclude it is broken — long before they notice anything about the quality of its answers.
This piece will break the round trip into its components, show where the time actually goes, and argue that for most deployments SBC placement moves the number more than the choice of model does.
What this article will cover
- The end-to-end latency budget, component by component.
- Why media path and SBC placement dominate the total.
- Barge-in and endpointing: the settings that decide perceived speed.
- What we hold ourselves to, and how it is monitored in production.