https://vapi.ai/ logo
Join Discord
Powered by
Assistant overly sensitive to background noise (in...
# support
m
Hi team, We’ve been using Vapi and have had a great experience so far. As we look to go deeper with the platform, we wanted to flag an issue we’ve been encountering. Our assistant appears to be overly sensitive to low-level background noise (such as light nearby conversation), which is causing frequent and premature interruptions (assistant stops speaking), even with denoising enabled and conservative interruption thresholds. Would appreciate any guidance on whether this could be due to configuration or tuning, and how we might address it. If helpful, we’re open to a quick 5–10 minute call as well. Thanks in advance!
Sample call ID which shows this issue: 019d86cc-288c-7000-be86-7990473543c2. Org ID here is: e5fb098a-cbb6-4a5b-8c21-dbc5b66f201c and assistant ID is 3ac19c57-3c05-4508-9e20-e650ac3142f2
s
We're investigating this and will soon get back to you.
m
Okay, appreciate it. We do have live customers who are a bit concerned here, so it would be super helpful to get your support here.
@Shaunak Just checking in here again - we're getting a lot of customer complaints here and would be good to have a resolution here as soon as possible.
s
Hi, We pulled the logs from call 019d86cc-288c-7000-be86-7990473543c2. The assistant (Emily) was getting cut off mid-sentence multiple times due to background audio triggering the voice detector - we can even see you saying "you cut off, can you repeat?" twice in the transcript. Why it happened Your assistant's stopSpeakingPlan is currently configured as:
Copy code
json
"stopSpeakingPlan": {
  "numWords": 0,
  "voiceSeconds": 0.5
}
numWords: 0 means the assistant stops speaking on any sound, without waiting for transcribed words voiceSeconds: 0.5 means just 500ms of audio is enough to cut it off In a noisy environment with background conversation, this fires constantly. The denoiser helps with static but can't always distinguish background speech from the caller speaking directly.
How to fix it Step 1 - Set numWords to 3 (most impactful change) This tells the assistant not to stop speaking unless at least 3 words are actually transcribed. Background noise rarely produces clean transcriptions, so this acts as a reliable second filter.
Copy code
json
"stopSpeakingPlan": {
  "numWords": 3,
  "voiceSeconds": 0.2,
  "backoffSeconds": 1
}
Step 2 - Raise the transcription confidence threshold Your transcriber currently accepts results at 40% confidence or higher. We saw noisy transcriptions at 0.55 confidence triggering the pipeline. Raise this to:
Copy code
json
"confidenceThreshold": 0.6
Step 3 (optional) - Enable Fourier denoising as a second layer
Copy code
json
"backgroundSpeechDenoisingPlan": {
  "smartDenoisingPlan": { "enabled": true },
  "fourierDenoisingPlan": {
    "enabled": true,
    "mediaDetectionEnabled": true,
    "staticThreshold": -35,
    "baselineOffsetDb": -15,
    "windowSizeMs": 3000,
    "baselinePercentile": 85
  }
}
Applying the changes via API
Copy code
bash
curl -X PATCH https://api.vapi.ai/assistant/b035c8d1-ff8a-4320-a2a0-7b941f238979 \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "stopSpeakingPlan": {
      "numWords": 3,
      "voiceSeconds": 0.2,
      "backoffSeconds": 1
    },
    "transcriber": {
      "confidenceThreshold": 0.6
    }
  }'
After applying, make a test call in your noisy environment - Emily should now complete full sentences without being cut off by background audio.
m
s
Thanks for confirming! In that case, could you try the other steps mentioned above?
m
@Shaunak the actual fix was changing the value for 
numWordsToInterruptAssistant
from 0 to 3. This is not exposed on the UI, and there is no documentation for it either - would urge you guys to update your docs. The other things you had suggested were already set to the right values.
s
Thank you so much for following up and sharing that fix. Since your setup is working now, I'll go ahead and close this thread, but if you face any further issues or need help with any other configurations, please feel free to open a new ticket.
PreviousNext
Powered by