I'm running into an issue where my assistant isn't hearing the caller. The call connects, the agent speaks the first message, but then the caller's audio isn't picked up, and the call eventually drops with a silence-timed-out reason. My Setup: Primary STT: Google (gemini-2.5-flash, Multilingual) — Required because we need multi-language support. Fallback STT: Deepgram (nova-2, en) What I'm seeing in the logs: Google STT initializes perfectly and appears to receive the audio stream (the logs show it clearing an audio buffer of 1172 chunks). However, it processes 0 words and generates 0 prompt/completion tokens. Because the Google API isn't returning a hard error (like a 500 or auth failure), VAPI seems to interpret this as actual silence. As a result, the Deepgram fallback never triggers, and the call just dies on the silence timeout. My questions: Is there a known issue or limitation with using gemini-2.5-flash as a real-time transcriber in VAPI? Since the fallback plan only seems to trigger on API errors, is there a way to configure a fallback if the primary STT returns empty transcripts for a certain duration? Thanks for your help!

Currently, Vapi's fallback plan only triggers if the primary transcriber fails with an API error—not when it returns empty transcripts or silence. There is no built-in way to trigger fallback based on empty results. Here's an example of how to configure a fallback plan (for API errors only):

json
"transcriber": {
  "provider": "google",
  "model": "gemini-2.5-flash",
  "language": "Multilingual",
  "fallbackPlan": {
    "transcribers": [
      {
        "provider": "deepgram",
        "model": "nova-2",
        "language": "en"
      }
    ]
  }
}

For now, consider monitoring for silence timeouts and handling them at the application level if you need more advanced fallback logic. See the [Vapi fallback documentation](https://docs.vapi.ai/customization/transcriber-fallback-plan) for details. Source: - [Transcriber fallback configuration](https://docs.vapi.ai/customization/transcriber-fallback-plan)

Call Flow: 1) AI say 1st message 2) Human respond to AI 3) STT do not return nothing NO ERR NO TEXT to LLM !

Hi,

gemini-2.5-flash

is not optimized for real-time streaming transcription. It receives the audio successfully but returns empty transcripts rather than a hard error, which is why Vapi interprets it as genuine silence and the Deepgram fallback never triggers. To answer your questions directly: 1. Yes,

gemini-2.5-flash

has a known limitation for real-time STT use cases - it's not suitable as a primary transcriber for live calls. 2. The fallback currently only triggers on hard API errors, not on empty transcript responses. We've noted this as a gap to address. The fix: Switch your primary STT to Deepgram Nova-2 with language set to

multi

- this gives you multilingual support and uses proper real-time streaming. You can keep Deepgram

nova-2 en

as your fallback or remove it entirely. Let us know if you need help updating the config! Best, Vapi Support