Thanks — I’m responding to your question about whether

session.say()

calls inside function tools can trigger the

e2e_latency

metric the same way normal assistant turns do. Short answer:

e2e_latency

is recorded on assistant

ChatMessage

items (per-turn metrics).

session.say()

does create speech and emits speech events (it returns a

SpeechHandle

and fires

speech_created

), but it does not automatically attach the same

e2e_latency

metric that appears on assistant turns. In other words,

e2e_latency

is not recorded for

session.say()

in the same way as a normal assistant turn. If you want latency for speech generated inside a function/tool, you can: • Measure timings around the function call yourself (record start time before invoking the tool and end time when the

SpeechHandle

reports audio ready or on the

speech_created

event). • Use the

SpeechHandle

lifecycle events to capture TTS timing (time to first byte / ready) and combine with your own measure of upstream LLM time if needed. Would you like a short code example for measuring those timings with

SpeechHandle

or guidance on where to hook the

speech_created

event? Sources: Understand and Improve Agent Latency | LiveKit | Agent speech and audio | LiveKit Documentation | Build Your First AI Voice Agent in Python: Complete Tutorial | LiveKit | Audio customization | LiveKit Documentation | The ReAct Pattern for Voice Agents and How AI Agents Think, Act, and Respon..., with interactive elements