We are currently facing an issue with the retell voice agent conversation flow where the agent suddenly starts talking in gibberish for no reason. Are you guys aware of this issue and what is the recommended solution to fix this? Please help
It looks like this may have been a transient issue with our voice service during that call. If you’re still experiencing this on more recent calls, please share those call IDs and we’ll be happy to investigate further.
I’ve getting the same issue. It is happening on calls with GPT 5.4 nano, claude sonet 4.5, claude opus 4.6. I haven’t tried other models but it seems to be a voice issue rather than an LLM issue.
We are still facing the issue where the agent suddenly starts talking in gibberish. Here are the details. We are really annoyed by this issue. Please resolve it soon.
Hello Mark1, I believe my team might have done that. Any change you can still diagnose the issue for us? Because, the current system also has this issue
The issue is in your conversation flow’s New_Patient_Collect node. The prompt uses bracket notation [read callingNumber as XXX-XXX-XXXX] as a formatting hint for the AI. When the caller had no caller ID (anonymous), the phone number variable was empty, but the model still attempted to follow the confirmation path and spoke the bracket text literally — which is what produced the gibberish.
Here’s how to fix it:
Replace the bracket notation with explicit natural-language instructions, e.g.: “Read the caller’s number back to them one digit at a time in the format XXX-XXX-XXXX.”
Add a hard guard for anonymous callers: make the prompt explicitly check if callingNumber is empty before offering to confirm digits, and skip straight to asking for their number manually if so.
Optionally, use Retell’s {{variable}} syntax for dynamic values instead of relying on the LLM to interpret bracket directives.
This is a prompt design issue — no platform bug. The TTS, LLM, and telephony layers all worked correctly; the model just output the formatting instruction as spoken text instead of interpreting it.