Hi Vapi Team, I’m currently facing an issue where the AI interviewer interrupts candidates while they are still speaking during an interview session. Many of our candidates are from the Philippines, and they tend to speak English slowly. They also often pause briefly or use thinking fillers such as “ahhh”, “umm”, or short pauses while formulating their answers. However, the AI frequently assumes they have finished speaking and starts responding, which interrupts them mid-answer. This significantly impacts the interview experience and makes it difficult for candidates to complete their responses naturally. I have already tried adjusting several parameters provided by Vapi, including: *

startSpeakingPlan.waitSeconds

*

smartEndpointingPlan.waitFunction

*

stopSpeakingPlan.voiceSeconds

*

stopSpeakingPlan.backoffSeconds

* Deepgram endpointing configuration Despite these changes, the issue still occurs. Below is the configuration I am currently using: startSpeakingPlan: { waitSeconds: 2.5, smartEndpointingPlan: { provider: "livekit", waitFunction: "700 + 4000 * max(0, x-0.5)" } }, stopSpeakingPlan: { numWords: 0, voiceSeconds: 0.2, backoffSeconds: 1.2 }, voice: { model: "sonic-2", voiceId: "b1a8a103-97c3-4deb-8010-02aedf2c63c9", language: "en", provider: "cartesia", generationConfig: { speed: 0.85, experimental: {} }, experimentalControls: { speed: 0.5, emotion: [ "positivity:high", "curiosity:low" ] } }, transcriber: { model: "nova-3", keyterm: ["Vishal", "Maurya"], keywords: ["Vishal", "Maurya"], language: "en-US", provider: "deepgram", endpointing: 350, confidenceThreshold: 0.5, } ``` I will also share some call IDs below so your team can review the logs. call id : 019cd7f3-5b0d-7115-b96f-2e600dd958e5 (timestamp 01:08) Call Id : 019cd6f3-db4c-7bba-abb9-f9784641fbc1 Thank you for your help.

To further reduce interruptions for slow or hesitant speakers, try making your LiveKit smart endpointing even more conservative. Update your configuration as follows:

json
smartEndpointingPlan: {
  provider: "livekit",
  waitFunction: "2000 + 6000 * x"
}

This will make the assistant wait longer before responding, especially when it's unsure if the candidate has finished speaking. Test this adjustment and see if it improves the interview flow for your candidates. For more details, see the [speech configuration documentation](https://docs.vapi.ai/assistants/speech-configuration). Source: - [Speech configuration documentation](https://docs.vapi.ai/assistants/speech-configuration)

Hi Vishal, Thanks for sharing the call IDs and configuration — we reviewed the logs. The main issue is that endpointing is triggering too quickly for slower speakers. In your case, the transcriber is finalizing transcripts after about 350ms of silence, and the smart endpointing wait function then allows the assistant to respond after ~700ms, which is too aggressive for candidates who pause or think aloud. To improve this, we recommend making the endpointing pipeline more conservative. The most effective changes are: • Increase Deepgram endpointing (e.g.

endpointing: 800

) so transcripts are finalized later. • Raise the wait function floor so the assistant waits longer before responding (e.g.

1500 + 3000 * max(0, x-0.3)

). • Increase

waitSeconds

to around 3.5 seconds so the AI waits longer after the user’s last voice activity. These changes should give candidates more time to pause and continue speaking without the AI assuming the turn is finished. Let us know if you’d like help testing these settings. Best, Vapi Support