LiveKit smart endpointing triggered on Deepgram partial before endpointing: 500 silence elapsed Vapi AI #support

LiveKit smart endpointing triggered on Deepgram pa...

N3squik

06/08/2026, 8:08 AM

Hi Vapi team, I’m investigating a premature assistant handoff where our Navigator assistant transferred after hearing only Hello., even though the caller/voicemail continued speaking shortly after. Call ID: 019e8424-c72f-7002-b292-b53a55cfeb50 Assistant config relevant parts:

Copy code

"transcriber": {
    "language": "en",
    "model": "nova-3",
    "provider": "deepgram",
    "endpointing": 500
  },
  "startSpeakingPlan": {
    "waitSeconds": 0.1,
    "smartEndpointingPlan": {
      "provider": "livekit",
      "waitFunction": "50 + 450 * x^3"
    }
  }

From the call log, the sequence appears to be: +4.247s Deepgram partial transcript: "Hello." +4.252s LiveKit prediction completed: 0.4680, awaiting 117.76ms +4.369s Endpointing triggered, provider: livekit +4.370s LLM request started +4.908s LLM tool call received +4.917s handoff initiated The confusing part is that the Deepgram transcript was logged as a partial transcript, not final, and the customer audio continued shortly after. Based on the local audio, the gap after “Hello” was roughly ~300ms before the next customer speech, so I expected Deepgram’s endpointing: 500 to prevent the turn from being finalized and sent to the LLM. Can you please explain why the Deepgram's endpoint did not work here?

Chiranjeet Mishra

06/08/2026, 6:55 PM

I am looking into it

Chiranjeet Mishra

06/10/2026, 7:32 PM

Hi, Can you help me understand whether you're seeing this behavior only with the configuration you shared, or does it happen with other configurations as well? Looking at your current setup, the endpointing configuration is extremely aggressive and optimized for very fast responses. This can sometimes cause interruptions or unexpected conversational behavior. As a test, could you try switching to a more balanced configuration, such as:

Copy code

{
  "startSpeakingPlan": {
    "waitSeconds": 0.6,
    "smartEndpointingPlan": {
      "provider": "livekit",
      "waitFunction": "700 + 4000 * max(0, x-0.5)"
    }
  },
  "stopSpeakingPlan": {
    "numWords": 1,
    "voiceSeconds": 0.4,
    "backoffSeconds": 2
  }
}

Then let us know whether the issue still occurs. This will help determine whether the behavior is related to the current endpointing settings or something else in the configuration. Thanks, Vapi Support Team

Previous Next