https://vapi.ai/ logo
Join Discord
Powered by
one word answers are not detected
# support
e
I notice that one word answers are not caught in some calls. For example, when user says kickball or no, it is not being detected in this call: 019d68e3-c072-7550-9bea-f2396c48e476. Furthermore, we see a lot of mistranscrption issues when user only speakd 2-3 word answers, we use these parameters. How can we improve mistranscription in these cases? • eotThreshold: 0.7 • eotTimeoutMs: 5000 • confidenceThreshold: 0.2 • waitSeconds: 0.2 • smartEndpointing: disabled
s
Hi, thanks for reaching out.
We're looking into this and will soon get back to you on this.
Hi, Thanks for sharing those details. Here are a few things you can try on your end to improve short utterance detection:
1. Lower confidenceThreshold Try dropping it from 0.2 to 0.1. This makes the transcriber less aggressive about filtering out low-confidence short responses. 2. Enable native endpointing Add endpointingMs: 250 alongside your current config. This can help catch short utterances that get cut off before they're fully processed. 3. Prompt for multi-word responses Where possible, design your prompts to encourage slightly longer answers. For example, instead of expecting "no", prompt for "no I don't" or "yes I do". This reduces reliance on single-word detection.
4. Add confirmation patterns For critical yes/no moments, have the assistant repeat back what it heard - e.g. "Just to confirm, you said no?" This acts as a fallback for cases where short answers may be misheard.
Give these a try and let us know if you're still seeing issues - feel free to share additional call IDs if needed.
e
thank you for your response @User , we use flux model from deepgram, so I guess we cannot change endpointing with this model?
s
Hi, endpointing is configurable with the Flux model! Flux has its own native end-of-turn detection, and since you've already disabled smartEndpointing, here's what to tune for better short utterance detection:
Copy code
{
  "transcriber": {
    "provider": "deepgram",
    "model": "flux",
    "eotThreshold": 0.5,
    "eotTimeoutMs": 3000,
    "confidenceThreshold": 0.1
  }
}
What each change does: - eotThreshold: 0.5 — makes end-of-turn detection more sensitive to short pauses - eotTimeoutMs: 3000 — reduces the max wait time before forcing end-of-turn - confidenceThreshold: 0.1 — reduces the chance of short words like "no" or "kickball" being silently dropped by Vapi's hallucination filter Give that a try and let us know if short utterances are being picked up!
PreviousNext
Powered by