Are we able to modify transcription confidence thresholds for acceptable transcriptions from the STT node?

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

i’m not interested in end of turn capabilities. I’m trying to understand if I can get the STT node to pass on transcriptions that were low confidence.

Thanks — I’m responding to your clarified question about whether the STT node can be configured to pass on (or filter out) low‑confidence transcriptions. Short answer: The STT node generally passes through the transcriptions the provider returns, including low‑confidence results. Confidence-related settings exposed by the STT node are primarily for turn/end‑of‑utterance and VAD behavior, not for filtering individual low‑confidence words or entire transcripts. If you need to accept or reject transcripts based on confidence, you’ll need to do that after the STT node (post‑processing), or rely on provider features if the provider exposes per‑word confidence that you can act on. What you can do next: • Check if your STT provider returns per‑word or per‑segment confidence scores. If it does, implement a small post‑processing step that: a. Inspects confidence values on words/segments. b. Drops or flags words below your threshold, or discards the entire segment if average confidence is too low. • If your provider exposes any confidence-related options (some vendors allow extra flags), you can use those to influence the returned results — but those options are provider‑specific and usually don’t automatically filter results inside the STT node. • If you tell me which STT provider you use (e.g., Deepgram, AssemblyAI, Speechmatics), I can: ◦ Confirm whether that provider returns per‑word confidence scores, and ◦ Give a short example of how to implement simple post‑processing to filter by confidence. Which provider are you using and do you want examples for dropping low‑confidence words or discarding whole segments?

Soniox. interested in dropping low confidence words.

Thanks — I’m responding to your request about dropping low‑confidence words from Soniox transcriptions. Short answer: Yes. Soniox returns per‑word confidence scores and you can drop low‑confidence words by overriding the STT node in your agent, inspecting the word-level confidences on each transcription event, and emitting a modified transcript with low‑confidence words removed. What to do (high-level) • Override your agent’s

stt_node

hook so you intercept transcription events as they stream from Soniox. • For each transcription event (interim or final), read the words/tokens and their confidence values. • Build a filtered transcript by excluding words below your chosen threshold (for example 0.5). • Replace the event’s text/alternatives with the filtered text before yielding the event so downstream components only see the cleaned transcript. Example concept (Python-style pseudocode) • The snippet below demonstrates the pattern: call the base implementation to get events, filter words by

word.confidence

, update the

event.alternatives[0].text

, then yield the event.

async def stt_node(self, audio: AsyncIterable[rtc.AudioFrame], model_settings: ModelSettings):
    # Stream transcription events from the default STT processing
    async for event in super().stt_node(audio, model_settings):
        # Only process transcript events that include word-level data
        if event.type in (stt.SpeechEventType.FINAL_TRANSCRIPT, stt.SpeechEventType.INTERIM_TRANSCRIPT):
            alt = event.alternatives[0] if event.alternatives else None
            if alt and hasattr(alt, "words"):
                threshold = 0.5  # set your confidence threshold here
                filtered_words = [
                    w.text for w in alt.words
                    if getattr(w, "confidence", 1.0) >= threshold
                ]
                # Join remaining words; you can preserve spacing/punctuation as needed
                alt.text = " ".join(filtered_words)
        yield event

Notes and tips • Choose your threshold based on testing. 0.5 is a common starting point but you may prefer higher for cleaner transcripts. • If Soniox sometimes omits per‑word confidence for certain tokens, guard with a default confidence (as shown) or treat missing values as low confidence. • Decide whether to apply filtering to interim results, final results, or both. Filtering interim results may make streaming feel choppy; many choose to filter only final transcripts. • If you want to discard entire segments instead of removing words, compute an average or minimum confidence across the segment and skip yielding the event when it’s below your cutoff. • Keep downstream consumers in mind: removing words can change timestamps or semantics. If you rely on precise timing, you may prefer flagging low‑confidence words rather than removing them. If you’d like, I can: • Provide a more concrete code example in the exact language/framework you’re using. • Show an example that flags low‑confidence words (e.g., wraps them in markers) instead of dropping them. • Suggest a simple test plan to pick an appropriate threshold. Which would you prefer? Sources: Soniox STT plugin guide | LiveKit Documentation | Pipeline nodes and hooks | LiveKit Documentation