What is turn latency?

Human conversation runs on roughly 200 ms gaps. People tolerate more from a phone line, but the tolerance is not large: past about 1.2 seconds a caller assumes the line dropped and starts talking again, which then collides with the agent's reply.

The total is a sum, not a single vendor's number: endpointing (deciding the caller has actually stopped), speech-to-text, the language model's first token, text-to-speech first audio, and network transport in both directions. A fast model behind slow endpointing still feels slow.

Because it is a sum, the only meaningful measurement is at the wire (recorded on the actual call), not from an API benchmark. Carrier recordings are also unreliable for this: they are normalised and re-timed, so they can hide or invent hundreds of milliseconds.