Hi, We're using Deepgram's Flux model for transcription and noticed that changing the endpointing value in our configuration directly impacts latency. According to Deepgram's documentation/support, endpointing should not affect behavior when using the Flux model. However, in our testing, adjusting the endpointing value (e.g., from the default to a higher/lower value) produces measurable changes in latency. Could you clarify: Is endpointing expected to be ignored when using the Flux model, or does it still apply?

Good catch but the latency changes you're seeing aren't coming from the old endpointing param. That parameter doesn't apply to Flux at all (Flux runs on /v2/listen, not /v1). What's actually affecting your latency are Flux's own turn detection parameters: eot_threshold lower = faster responses, higher = fewer cutoffs eot timeout ms tweak this if users speak with longer pauses One thing to check on VAPI: Make sure smartEndpointingPlan is set to Off otherwise it'll override Flux's built in detection and cause exactly the kind of inconsistent latency you're describing.

@Kashan @ UnitZero AI is correct here. because deepgram flux uses its own endpointing mechanism, using VAPI

smartEndpointing

will conflict with flux. this is the primary reason why we advise to disable

smartEndpointing

on VAPI side