(using agents version 1.2.8)
Hello,
Problem: Trying to measure total latency (from end of user speech to start of agent speech) in 2 pipelines (1: realtime API, 2: separate STT, LLM, TTS)
What I've done:
For pipeline 1: attempted to find total delay by summing RealtimeModelMetrics.ttft + EOUMetrics.end_of_utterance_delay (
docs say turn_detection must be set to "vad" for EOU delay to be captured, and I have done this). However, EOUMetrics.end_of_utterance_delay is still not showing up in my logs.
I also tried to use EnglishModel() instead of VAD as turn_detection, but EOU metrics still did not show up.
For reference, this works in pipeline 2:
calculating total delay by EOUMetrics.end_of_utterance_delay + LLMMetrics.ttft + TTSMetrics.ttfb, i.e., all the metrics are received as expected
Questions:
1. How can I find EOU delay in Realtime pipeline (I've already set turn detection to vad)?
2. Could you please provide more detail on these 2 parameters of RealtimeModelMetrics:
duration: float
"""The duration of the response from created to done in seconds."""
ttft: float
"""Time to first audio token in seconds. -1 if no audio token was sent."""
For calculating total delay, should "duration", or "ttft" be used? (eou delay + duration OR eou delay + ttft)
Thanks in advance!!
cc
@refined-appointment-81829