019a7bbc-94ce-7885-898a-53c2c60e79c8

Suggestion: Check if Vapi's endpointing is correctly detecting when the user has finished speaking, as this controls when the LLM is called. You can review your endpointing configuration or logs to ensure the system is not ending the turn prematurely. Example:

// Ensure endpointing is enabled and configured
// (Pseudo-config, adjust for your setup)
{
  "endpointing": {
    "enabled": true
  }
}

If endpointing is misconfigured, the LLM may not be triggered as expected. See more in the orchestration/endpointing section of the documentation. Source: - [How Vapi Works](https://docs.vapi.ai/how-vapi-works)

@Sahil hey team, sorry to ping you directly, this is a prod customer, and almost all the call are failing

Try these settings for deepgram flux:

{
  "transcriber": {
    "provider": "deepgram",
    "model": "flux-general-en",
    "eotThreshold": 0.5,        // Lower from default 0.7 for better responsiveness
    "eagerEotThreshold": 0.35,  // Helps prepare responses earlier
    "eotTimeoutMs": 2500,       // Force EOT after 2.5s of silence
    "language": "en"
  }
}

adjusting these value, what difference would make?

We reviewed the logs for the latest call ID you shared. The issue wasn’t a timeout failure — it was caused by the transcriber’s end-of-turn confidence being slightly too low to be accepted. What actually happened - Flux did send an

end_of_turn

signal at

22:59:46

- But its confidence was 0.6875, just below the configured threshold - Flux kept streaming partial transcripts afterward - Each partial update reset the silence timer - The system stayed stuck in agent_turn, waiting for a stronger signal - Later user speech was detected but ignored because the agent was still speaking Why this caused the problem Because confidence didn’t reach the 0.7–0.8 expected range, the system didn’t accept the turn ending, so it never handed control back to the LLM. Suggested settings to fix this

json
{
  "transcriber": {
    "eotThreshold": 0.65,         // Accept the 0.6875 signal
    "eotTimeoutMs": 3000          // Force turn closure after silence
}

In short That 0.6875 confidence was just below your threshold, so the turn never ended and Flux kept streaming partials indefinitely.

019aa913-07cb-7229-81f3-2b007c1e3ce5 @kyle what about this one?

If the transcripted speech is below the confidence threshold, it will ignore the speech. In the call id

019aa913-07cb-7229-81f3-2b007c1e3ce5

the confidence for the speech at the end of the call was rated 0.394 which is below the 0.4 confidenceThreshold.

🔵 01:03:36:814 Ignoring low confidence transcript (0.394 < 0.4) Yeah. I understand. Like, for your assistance. But I'm at it. Correct. So we'll go through the paperwork, uh, find out where I'm going, uh, go through my height requirements, like, you know, there are height requirements, height restrictions, road restrictions. So I will follow that, uh, according to my length and weight of the truck, and I will plan how many kilometers I need to travel, uh, before I can start my break. Uh, more particularly, like, I have on BFM. So, like, you know, for first five hours, I will drive, um, I So after getting the paperwork, I will go go through the truck, do a pre start inspection, go around the truck, check all the lights, uh, indicators, and, uh, and check the tires are okay. Um, the weight is properly loaded. Uh, if it is not, we'll get it, uh, done, verify the hook hook up of the trailers are, uh, perfectly fine for the road. And once it is done, I will start this, uh, truck up. I will inspect that it's good, like, all the oil labels and all the truck warnings are okay, uh, to go ahead. After doing that, like, you know, my work has already started, so I will

@kyle 019aa913-07cb-7229-81f3-2b007c1e3ce5 -> Yes it was below, however shouldn't vapi automatically extend the timeout? As the call has been ended by VAPI meanwhile the candidate was speaking

Web calls have faster transport than various phone service providers