how to get transcript confidence for chat message

Thanks for reaching out, Aman. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

i'm using deepgram as stt

Thanks — that’s helpful. You said you’re using Deepgram for STT; I’ll explain how to get the transcript confidence and show short examples you can adapt. What to expect • Deepgram returns a confidence score with each transcript alternative. The score is a number from 0 to 1 (higher = more confident). • When using Deepgram with LiveKit, the confidence is included in the transcription events (for example via the

user_input_transcribed

event or when processing the

SpeechStream

from the Deepgram plugin). You should read the

alternatives

array on the event and extract the

confidence

from the chosen alternative. Two concise examples you can use 1) Listening to LiveKit transcription events (Python-style pseudocode) • This shows extracting confidence from the event payload when LiveKit emits

user_input_transcribed

.

@session.on("user_input_transcribed")
def on_transcript(transcript):
    # Only proceed if we have alternatives
    if transcript.alternatives:
        # Choose the primary alternative (index 0)
        alt = transcript.alternatives[0]
        text = alt.text
        confidence = alt.confidence
        print(f"Transcript: {text}  Confidence: {confidence}")

2) Handling Deepgram SpeechStream directly (async streaming) • If you are consuming the Deepgram stream yourself, check each event for alternatives and confidence.

async def recognize_task(stream: deepgram.SpeechStream):
    async for event in stream:
        if event.alternatives:
            alt = event.alternatives[0]
            final_text = alt.text
            confidence = alt.confidence
            if event.is_final:
                print(f"Final transcript: {final_text} (confidence={confidence})")
            else:
                print(f"Interim transcript: {final_text} (confidence={confidence})")

Notes and tips • Alternatives: Many responses contain multiple alternatives. Use whichever alternative makes sense (commonly the first). • Interim vs final: You’ll receive interim events with confidence as well — typically use the final event for production decisions. • Thresholding: If you want to filter low-confidence transcripts, compare the confidence to your threshold (for example, 0.7) before creating a chat message or taking an action. • End-of-turn logic: Deepgram and LiveKit may use confidence thresholds (e.g.,

eager_eot_threshold

or

end_of_turn_confidence_threshold

) internally to decide end-of-turn; those are separate and configurable if you need different behavior. If you want, tell me: • Whether you want sample code for another language (Node, Go, etc.), or • Where you’re consuming transcripts (event handler vs plugin callback) and I’ll give a tailored snippet for that setup. Sources: Deepgram STT plugin guide | LiveKit Documentation | An open source stack for real-time multimodal AI | An open source stack for real-time multimodal AI