Yes, this issue has been reported. From Thursday, December 18, 2025, there’s a documented case where an agent verbally outputs tool-related technical data during calls(1).
The issue you’re experiencing—where the agent reads JSON extracts aloud—appears related to how the LLM handles function calling responses. Here are potential solutions:
1. Configure Speech Behavior for Extract Dynamic Variables
When setting up your Extract Dynamic Variables tool, ensure “Speak during execution” is disabled(2). This prevents the agent from speaking while the extraction function executes(2).
2. Update Your Prompt
Add explicit instructions to your prompt about when to invoke the extraction function(2):
When user states his name and phone number, please extract the information by calling the `extract_user_details` function.
3. Consider LLM Provider
One documented case suggests that Gemini models don’t always perform well with tool calls, and recommends using GPT 4.1 models instead(1).
4. Review Speech Configuration
Check your agent’s “Speak after execution” setting for the extraction tool(3). This controls whether the agent continues talking after the function completes(3). You may want this disabled to prevent the agent from verbalizing extraction results(3).
5. Normalize for Speech
Ensure normalize_for_speech is set to true in your agent configuration(4)(5)(6). This normalizes text before speech synthesis, which can help prevent raw technical data from being spoken(4)(5)(6).
If these steps don’t resolve the issue, you may want to reach out through official support channels:
- Visit Contact Sales
- Ask questions on Discord
- Visit https://github.com/RetellAI/