HI Vapi team, we're experiencing an annoying silent issue making the call experience terrible. Seems some fails in the ElevenLabs Text-to-Speech API request is not triggering the fallback defined in the

fallbackPlan.

Expected behavior: When the TTS request fails, the system should automatically activate the fallback plan as configured Actual behavior: The fallback plan was not triggered, and the call was created but there are no audio from assistant, e.g first message. Impact: Unresponsive assistant causing a terrible call experience. Vapi CallID: 019dac80-5324-7550-9df7-f17c25bd1127 (sharing just one for simplicity, but we have a few ones) There are no error in the Vapi logs, which is a little bit weird. https://cdn.discordapp.com/attachments/1495908712542638200/1495908714866016368/image.png?ex=69e7f50e&is=69e6a38e&hm=03ae6e74982a9bf79f2442e8573d15a168646bf1180d4606ac75dcf5b41b820d&

Suggestion: Please verify that your assistant's voice configuration includes a properly structured

fallbackPlan

as shown below. The fallback voices should be listed in the

voices

array, and the configuration must be valid JSON:

json
"voice": {
  "provider": "eleven-labs",
  "voiceId": "your-elevenlabs-voice-id",
  "fallbackPlan": {
    "voices": [
      {
        "provider": "azure",
        "voiceId": "andrew"
      }
    ]
  }
}

Ensure the fallback provider and voiceId are correct, and test by simulating a failure to confirm the fallback triggers as expected. For more details, see the [Vapi voice fallback documentation](https://docs.vapi.ai/voice-fallback-plan). Source: - [Voice fallback configuration](https://docs.vapi.ai/voice-fallback-plan)

Yes, it is.

Hi! I’ve dealt with similar silent TTS failures in Vapi setups, especially with ElevenLabs not properly triggering fallback flows. This usually comes down to how the error is surfaced (or not) in the TTS layer, so the fallbackPlan never actually gets invoked. I can help you trace where the failure is happening (request, response, or timeout) and enforce a reliable fallback trigger so your assistant never goes silent again. Happy to take a quick look at your config + logs and fix this fast. @Leo

Hey Adam. Yes, I agree with you.. Seems Vapi is not triggering the fallback properly.

Yes

We could apply a minor fix on our end, but the goal is to understand why Vapi fallback feature didn't work.

hey leo, apologies for the delayed response. we looked into the call id and found that your firstMessageMode is assistant-speaks-first-with-model-generated-message, which means the LLM is gating the first audio. with no first token and no completed response, no greeting text was ever generated to send to TTS, so the call sat silent. what you want to do is either fix the custom-llm implementation so that the model doesn't fail the first token generation + response, or use a preset model from our model selection menu. once you have the model returning responses and tokens, the TTS should kick in.

Thanks @kyle , we'll keep monitor.