Just wanted to confirm few things we are using deepgram stt whose trancripttion latency is 200 ms and we’re using Livekit’s multilingual turn detection model here we’ve setup min_delay as 400 ms and max_delay as 2 secs with this config my eou_delay is around ~ 900 ms so is it because of turn detection is taking more time then required or something else as i was under the assumption that turn detector take around 50-180 ms not this much time can shed some light on it
900ms is longer than I would expect, I usually see about half that.
As for what is ‘typical’, many developers compare the experience to speaking with the LiveKit agent hosted at https://livekit.com/ (select ‘Talk to LiveKit Agent’), and that will give you metrics as you speak.
I’m curious, what do you use to measure turn latency?
You measure turn latency using the per-turn metrics field on each ChatMessage. The key value is e2e_latency, which represents the time from when the user stopped speaking to when the agent began responding. These metrics are available via session.history or the conversation_item_added event.
For deeper analysis, you can also inspect component-level metrics such as end-of-utterance delay, LLM time-to-first-token (TTFT), and TTS time-to-first-byte (TTFB). All of this is described in the Per-turn latency section of Data hooks.