Gemini 3.1 Flash TTS: Volume fading / "whispering" style drift during long-form audio generation

[yooo1999]

Description of the bug:

When generating long-form speech (generations longer than 2 minutes) using the gemini-3.1-flash-tts-preview model, the generated audio suffers from severe volume fading and prosody drift.

Specifically, if the model is instructed to speak in soft or relaxed tones (either via inline style tags like [softly], [relaxation], or through system instructions), the autoregressive nature of the model causes a style/volume drift over time. By minute 2, the volume drops significantly, and by minute 4, the voice degenerates into a barely audible whisper or total silence.

To work around this, we had to add a flag to completely strip all inline style tags, which forces the model to generate in a flat, neutral tone. However, this defeats the purpose of Gemini's rich expressive audio capabilities.

Steps to Reproduce

1. Initiate a TTS request using the Python SDK with response_modalities=["AUDIO"].
2. Model: gemini-3.1-flash-tts-preview
3. Voice: Kore (or any other built-in voice).
4. Input Text: A long-form story (500+ words, approx. 4–5 minutes of speech) containing style directives in the system prompt or inline tags (e.g., [softly], [relaxation], [serenity]).
5. Observe the volume envelope of the output audio over time.

Actual vs expected behavior:

Expected Behavior

The voice should maintain a consistent, audible volume level throughout the entire length of the narration, adjusting the expressive style without dropping the master gain to zero.

Actual Behavior

• 0:00 - 1:30: High quality, expressive, correct volume.
• 1:30 - 3:00: Gradual but steady drop in volume.
• 3:00 - End: The voice degenerates into a quiet whisper, mumbling, or complete silence.

Any other information you'd like to share?

• SDK: google-genai (Python)
• Model: gemini-3.1-flash-tts-preview
• Output Format: AUDIO

[kkorpal]

Hi @yooo1999 Thanks for bringing this to us! I will bring this to our engineering team to investigate.

[kkorpal]

Hi @yooo1999, Could you try testing this with our other TTS preview models (like gemini-2.5-pro-tts-preview or gemini-2.5-flash-tts-preview) and let us know if you still experience the same volume drift?

[yooo1999]

I tested generating the exact same full story transcript (2,964 characters, ~560 words) in a single, un-chunked API request (no splitting, no inline audio/SFX tags) using both TTS models:

1. gemini-3.1-flash-tts-preview:
   • Total Audio Duration: 5:31 min
   • Result: Suffers from extreme volume drift / fade-out. Starting at ~3:26 min, the vocal volume gradually decays down to 0.82% of its initial amplitude (becoming completely inaudible / whispering by the end).
2. gemini-2.5-flash-preview-tts:
   • Total Audio Duration: 5:14 min
   • Result: No volume drift detected! The volume and clarity remain consistent throughout the entire 5+ minutes.
     This confirms the gradual volume decay / whispering bug is specifically introduced in gemini-3.1-flash-tts-preview when outputting long single-request audio (over 3 minutes).

[kkorpal]

Hi @yooo1999 , Thanks for confirming, I will discuss with the team.