Hey all, curious if anyone else is experiencing their call transcripts being truncated? I've been seeing that the last few seconds of my calls with assistants are not reflected in the transcript (via api or on the dashboard). For example, the entire last message and sign-off from my assistant will be said to me on the call, but in the transcript, the previous message is listed as the last one. This is especially problematic when a call doesn't reach the end of the flow, and I'm unsure how it ends. Basically, if it's said on the call by either party, I need it in the transcript. The transcriber I'm using is Deepgram Nova 3, with a confidence threshold of 0.4. Any help or insight you all could provide would be much appreciated. 🙂

Hi there! 👋 I understand how frustrating it can be when your call transcripts are missing important parts of the conversation. I'm unable to tell you if other users are specifically experiencing this same issue right now, but I can definitely help with the technical aspects of your transcript truncation problem! Based on the information I've found, there are several potential reasons why your call transcripts with Deepgram Nova 3 might be getting truncated: ## Endpointing Settings The default Deepgram endpointing value of 10ms may be too aggressive, causing premature cuts in transcription. According to [information I found](https://discord.com/channels/1211482211119796234/1343636482874933340), increasing the endpointing value from 10ms to 300ms can give the system more time to process the complete utterance and avoid premature interruptions. ## Wait Time Configuration Increasing the

waitSeconds

parameter in the

startSpeakingPlan

configuration to 0.8 can help reduce transcript truncation issues. This adjustment allows the system more time to capture the full context before responding, as mentioned in [several discussions](https://discord.com/channels/1211482211119796234/1337356648259387412). ## Confidence Threshold Impact Your confidence threshold of 0.4 could be contributing to the issue. According to [information I found](https://discord.com/channels/1211482211119796234/1333510134617604147), Vapi filters out words with confidence scores below the threshold, which can lead to incomplete transcripts. This filtering can cause entire replies to be omitted, especially when dealing with accents or foreign names. ## Configuration Adjustment Example You might want to try adjusting your configuration with something like this:

json
{
  "transcriber": {
    "provider": "deepgram",
    "model": "nova-3",
    "language": "en",
    "endpointing": 300
  },
  "startSpeakingPlan": {
    "smartEndpointingEnabled": true,
    "waitSeconds": 0.8
  }
}

Hey! To help track down this issue, could you share: - The call ID - When exactly this happened (the timestamp) - What response you expected to get - What response you actually got instead This would really help us figure out what went wrong!