Hi, I do often run into trouble with Transcriber-lag on VAPI dashboard, e.g. gemini sometimes responds in 1s (as expected) but sometimes jumps to >3s, which makes even the agent trigger "silence from user" flows. - and often just... does not transcribe. - is that related to over-load of the model via VAPI? (maybe own key would help?)