Why can a simple smart-glasses request feel slow?
Delay accumulates across the whole request. Speech detection, network calls, service access, model generation and display delivery can each add waiting. A fast model response is only part of a fast glasses experience.
Measure the moment the wearer gets something useful
Start the clock at the user's action and stop it when the first usable response appears or is spoken. An empty loading card is feedback, but it is not the answer. Separately recording when listening starts, recognition finishes and content appears makes it possible to identify the slow stage.
Suppose a model answers promptly but its result waits behind several old display updates. Replacing the model would not solve that delay. Giving new user requests priority and expiring stale cards could. This is why a timeline is more useful than a single model-speed number.