Deepgram Transcription Confidence Threshold
I have a case where Confidence Threshold may not have worked correctly.
The configuration is under Assistants, inside 'Transcriber' section, when 'Deepgram' is selected.
The description of the configuration in the UI is
I have an assistant using Deepgram Flux General English EN with a confidence Threshold of 0.5.
During a call, a transcript turn was given a confidence value of 0.2169
However, that doesn't seem to impact what was sent to the LLM
I'm expecting the LLM to know that the transcription 'The mature post' was low confidence, or at least filtered out.
Could you explain how the confidence configuration is meant to work?
The configuration is under Assistants, inside 'Transcriber' section, when 'Deepgram' is selected.
The description of the configuration in the UI is
Confidence Threshold
Transcripts with a confidence score below this threshold will be filtered out.
I have an assistant using Deepgram Flux General English EN with a confidence Threshold of 0.5.
During a call, a transcript turn was given a confidence value of 0.2169
{
"id": "log-56",
"level": 30,
"time": 1763194235360,
"timestamp": 1763194235360000000,
"severityText": "INFO",
"severityNumber": 9,
"body": "Deepgram flux final transcript",
"attributes": {
"category": "transcriber",
"callId": "019a8690-ab8e-7dd4-abc4-0399690c00bc",
"orgId": "5143d15b-8d5a-4df0-882d-daecf9ec66a1",
"provider": "deepgram",
"channel": "user",
"transcript": "The mature post.",
"confidence": 0.2169,
"numWords": 3,
"duration": 1486,
"latency": 479
}
}However, that doesn't seem to impact what was sent to the LLM
{
"id": "log-59",
"level": 30,
"time": 1763194235363,
"timestamp": 1763194235363000000,
"severityText": "INFO",
"severityNumber": 9,
"body": "Google HTTP request",
"attributes": {
"category": "model",
"callId": "019a8690-ab8e-7dd4-abc4-0399690c00bc",
"orgId": "5143d15b-8d5a-4df0-882d-daecf9ec66a1",
"provider": "google-aistudio",
"model": "gemini-2.5-flash",
"request": {
"model": "gemini-2.5-flash",
"messages": [
{/* truncated */},
{
"role": "user",
"parts": [
{
"text": "The mature post."
}
]
}
],
/* truncated */
}
}
}
}I'm expecting the LLM to know that the transcription 'The mature post' was low confidence, or at least filtered out.
Could you explain how the confidence configuration is meant to work?
