Hi,
We are currently evaluating google/gemini-3.1-flash-tts-preview for a multilingual production application.
The speech quality, naturalness, prosody and multilingual performance are excellent. However, we are experiencing one significant issue: the perceived voice identity can change noticeably between separate TTS requests, despite using the same selected voice and configuration.
One generation may sound exactly right, while the next can sound noticeably like a different speaker. The variation can affect timbre, pitch, perceived age and overall vocal character. A subsequent generation may then return closer to the original voice.
Natural variation in expression and prosody is desirable. The problem is specifically speaker identity consistency, not expressive variation.
We currently access the model through an OpenAI-compatible gateway rather than directly through the Gemini Developer API.
Could you please advise:
- Is voice-identity drift between separate requests a known issue or limitation of
gemini-3.1-flash-tts-preview? - Is there a recommended way to maintain a consistent speaker identity across consecutive TTS requests while preserving natural prosody?
- Are there parameters, prompting techniques or session/context mechanisms intended specifically for voice consistency?
- Would you recommend testing the model directly through the Gemini Developer API to rule out the gateway layer?
- If this is a current model limitation, is improved voice consistency planned for an upcoming version?
- Is there an approximate timeframe or preview model available for testing?
This is currently the main issue preventing us from moving the model into production, and we are hoping to make a production decision within the next few weeks.
We can provide anonymized audio examples and request details if needed.
Thank you.