Voice latency
The engineering that makes the other four capabilities fit inside a conversational turn.
The budget. A voice turn allows a few hundred milliseconds before the caller hears the pause and starts talking over the agent. Inside that: speech recognition, retrieval, generation, compliance check, speech synthesis. [Insert your measured per-stage breakdown here — do not ship this page with an estimated number.]
Why local wins. Every hop off your network is budget you cannot recover. Local inference is not an optimisation here; it is what makes the interaction feel like a conversation rather than a walkie-talkie.
What we do about it. Model selection sized to the budget rather than to a leaderboard, batching and GPU sharing across the capabilities that run on every turn, and streaming so synthesis starts before generation finishes.