I'm using a bring-your-own server setup. I can see in the logs that when I interrupt the agent speech my inquiry gets transcribed properly by Deepgram, however this message is not included in the messages sent in the next Custom LLM request. This leads the agent to respond to the initial query a second time. We've been getting complaints from users having to repeat themselves after an interruption.

Can you share your call id for which you are facing this issue?

019cc6fa-5062-7225-88a2-0a8cb69cab05 It's actually concatenating the interruption transcript with the previous transcript and sending them both back through as the last message. The partial agent message is not included in the message list. This causes the agent to believe it needs to answer the first half(the first question again) and essentially starts repeating itself.

Hi, Sorry for the delay. We do see the issue and our team is currently investigating it. Please allow us some time while we look into this. Thank you for your patience.

Okay, but we rely on custom-llm implementation..

Hi, Thank you for your patience, and sorry for the delay while I investigated the issue. For call 019cc6fa-5062-7225-88a2-0a8cb69cab05, the caller is interrupting the bot before it finishes speaking. When this happens, the bot’s context can get slightly out of sync, leading to responses that feel “one question behind.” ---- Recommended Fixes: • Enable backgroundDenoisingEnabled = true → Helps distinguish real speech from background noise • Set backchannelingEnabled = false → Prevents false interruption triggers • Adjust silenceTimeoutSeconds / endpointing → Use a slightly higher endpointing value (e.g., 300ms) to give the bot more time before processing new input • Update your system prompt → Instruct the model to focus only on the latest user question and ignore interrupted responses ---- How to Verify: 1. Update the assistant configuration with the above settings 2. Make a test call and interrupt the bot mid-response 3. Confirm the bot answers the most recent question correctly 4. Review the call transcript in the Vapi dashboard to ensure alignment between user queries and responses ---- Thank You

This isn't really addressing my issue. You mentioned the team was working on it; is there an update on that?

Hi, Could you please try switching the transcriber to Nova 3 and test the same scenario again? From what we’re seeing, this behavior seems to happen when the assistant gets interrupted mid-response, which may indicate a transcriber-side issue with interruption handling or partial transcript timing.