Are you using a custom LLM or one of Vapi's built-in providers? That affects how much of the latency is coming from the STT relay vs. the response pipeline.