latency benchmark lab

Replace a latency guess with 6 timestamps.

Enter production trace offsets, import JSON, or use the browser stopwatch during a test call. The result keeps each boundary separate so the slowest service is visible.

Stopwatch idle
speech end to playable audio760 ms

Slowest measured boundary: model first token at280 ms.

endpointing180 ms
final transcript80 ms
model first token280 ms
speech first byte150 ms
playback buffer70 ms

Use the same event definition in every provider test

Start the clock when caller speech actually ends. Measure endpoint decision, final transcript, first model token, first audio byte received by the runtime, and first audio that can be played to the caller.