in the last couple of weeks, we've been experiencing microphone cutoffs in our conversations. this a good call id as example

019cc452-c4be-7995-911e-0b8c848efded

if you have access to the recording, from

1:15

and going forward, user's voice seems to be breaking off which causes our AI either to interrumpt or reply not meaningfully. this issue has been killing our experience for a while already, please hope you can help us!

Hi, While the core fix is being addressed on our side, here are a few workarounds you can try in the meantime: 1. **Disable LiveKit endpointing**: Switch to server-side turn detection by setting

serverVad

as the turn detection method instead of using LiveKit endpointing. This uses a heuristic approach from the start and can behave more consistently. 2. **Increase silence threshold**: If you're using custom

serverVad

settings, try increasing

silenceDurationMs

(for example to 500–800 ms). This gives the system more time to confirm the speaker has finished before the assistant responds. 3. **Disable backchanneling**: Set

backchanneling: false

to reduce cases where the assistant interrupts during short pauses. 4. **Check network stability**: Since this is a

webCall

, audio quality depends on the caller’s connection. If users are on unstable Wi-Fi, switching to a wired connection or using a phone-based call (SIP/PSTN) can improve stability. Best regards, Oshi Raghav Customer Support Team | Vapi

ok, so there is an ongoing issue about this on your side? where can I follow the updates? I 'm also going to check those settings, thanks.

Hey, Sorry for the delay. If you are still facing this issue, it may be because your calls are still routing through LiveKit endpointing. If the endpointing provider hasn’t been explicitly changed, calls will continue hitting this path, which has a known instability. We recommend switching to transcriber-based endpointing and bypassing LiveKit entirely by using your transcriber’s built-in endpointing from Deepgram. Example configuration:

{
  "transcriber": {
    "provider": "deepgram",
    "model": "nova-2-general",
    "endpointing": 300
  },
  "startSpeakingPlan": {
    "smartEndpointingEnabled": false
  }
}

This uses Deepgram’s ML-based endpointing and avoids the request ID mismatch issue seen with LiveKit. Please try this configuration and let us know if the issue persists.

we made these changes and we're still getting similar issues, please inspect this call id 019ce5ea-0ad0-7002-8b67-88eb6ff96c78

Hi, sorry for the delay, i am looking into your issue.