STT Partials and transcripts
# support
f
For many transcripts it happens that the end-of-call-report transcript section misses users turns completely since they only are send as STT partials and the completes ones (final) are never appearing. I know that transcripts are filtering out partials but why are there no finals for some user turns?
v
The likely reason some user turns are missing final transcripts is that transcripts with low confidence (below the default threshold of 0.4) are discarded. To handle these cases, add the following hook to your assistant to prompt users when a low-confidence final is detected:
Copy code
json
{
  "hooks": [
    {
      "on": "assistant.transcriber.endpointedSpeechLowConfidence",
      "do": [
        { "type": "say", "exact": "I'm sorry, I didn't quite catch that. Could you please repeat?" }
      ]
    }
  ]
}
This ensures low-confidence user turns are acknowledged and gives users a chance to repeat themselves, improving transcript completeness. See documentation for more details: [Vapi Assistant Hooks](https://docs.vapi.ai/assistants/assistant-hooks) Source: - [Vapi Assistant Hooks](https://docs.vapi.ai/assistants/assistant-hooks)
f
The instances have threshold higher than 0.4 (i confirmed there is no "body":"Transcript ignored - low confidence") but are only send as partial no final messages in the raw call logs.
The original call log ID was: 019d1b18-0a48-7000-82fa-2c8fafcff133
a
Have you been able to solve this or you stilll need help? @Filippo
f
Hi Adam! Not solved.
a
Let’s connect @Filippo
f
I think i found the issue: I think SST final transcript is never send because the onNoPunctuationSeconds timeout fires first. What happens: Assistant asks question User says short answer (~400ms, short utterance) Deepgram emits a partial: "XXX" (conf 0.484, no punctuation) onNoPunctuationSeconds (1.5s) fires → new turn starts, TTS context resets Deepgram never produces a final for this utterance LLM is invoked with no new user message — the user's answer is lost This consistently affects short, low-confidence utterances like names and cities where Deepgram doesn't add punctuation. The partial exists but never gets promoted to a final because the turn already moved on.
There should be an option to wait for the final