Bug Report: gemini-3.1-flash-tts-preview cuts off audio after two words and freezes

Gemini API
,
1

I am experiencing a critical issue with the gemini-3.1-flash-tts-preview model inside the Google AI Studio interface. When trying to generate Portuguese (PT-BR) text-to-speech outputs using custom voice profiles and prosody instructions, the model consistently fails. It only speaks the first two words of the text, abruptly cuts off the audio generation stream, and completely freezes the UI.

Steps to reproduce:

  1. Open Google AI Studio and select the gemini-3.1-flash-tts-preview model.
  2. In the System Instructions / Scene field, paste a detailed custom voice configuration in Portuguese.
  3. Use a standard voice setting (e.g., Callisto or Kore) with a Temperature around 0.8.
  4. Input a paragraph of plain text containing inline prosody tags such as [whispering].
  5. Click “Run”.

Expected Behavior:
The model should generate the complete audio file for the entire paragraph, respecting the rhythm, pauses, and speech styles defined in the prompt.

Actual Behavior:
The model starts the audio delivery but stops abruptly after the first two words. The generation process gets stuck, and the interface freezes.

Environment details:

  • Model: models/gemini-3.1-flash-tts-preview
  • Language: Portuguese (Brazil - pt-BR)
  • Interface: Google AI Studio Web Playground