Title: Gemini 3.8 TTS / voice is a downgrade for personal-assistant and companion apps
I build two voice products on Gemini: a personal assistant and a companion app. Both run on the Aoede voice with carefully tuned prompts, and on the current Live model they sound natural, warm and like a real person.
When 3.8 was released I tested it on both. It failed:
Lack of personality. The voice is flat and generic. The warmth and character we have tuned for months are gone.
Shorter, clipped replies. It gives the minimum and stops, which kills a natural conversation.
For a PA or companion, personality and natural conversation are the product. A model tuned for short, neutral answers may suit call centres or readouts, but it is a step backwards for us.
I would like Google to:
Acknowledge that 3.8 is a regression for personal-assistant and companion use.
Keep the current Live model and voices available until there is a proper replacement.
Build, or offer a mode for, a PA/companion-friendly voice model that keeps personality, warmth and natural reply length.
Happy to share side-by-side samples if that helps.
I’m seeing the same thing. 3.8 seems technically better and the audio is smoother, but the conversational experience is a major regression.
I tested several voices and they all have the same problem: flat, monotonous and robotic delivery. The personality and natural reactions I get with 3.1 are gone. I also experienced unexpected language switching.
I really hope Google fixes this in 3.8 and brings back the natural, human conversational quality of 3.1.
I feel the same way—Gemini 3.1 Live’s generated speech is much more creative than Gemini 3.8 Live’s. You could test a quick example in Google AI Studio to compare them: use the system instruction ‘You are a singer, always speak with a melody’ and ask it to ‘sing something’ or ‘sing Happy Birthday.’ Gemini 3.1 Live is much more prone to singing with a clear melody, whereas 3.8 rarely captures it. The voice in 3.8 just sounds less interesting.
I’m seeing a similar issue. I use Gemini Live in a voice product, and 3.1 sounded much more natural, human and expressive. With 3.8, the voice feels flatter and less natural.
I’ve tried adjusting the voice direction in the prompt to make it more responsive and expressive, but I haven’t managed to recover the same quality. I’m wondering whether this is a model-level difference or whether there are prompting or configuration changes I’m missing.
I’ve also been experiencing increased latency with 3.1 recently, so switching back doesn’t fully solve the problem.
Has anyone found a practical way to get closer to 3.1’s naturalness on 3.8? Or a way to reduce latency on 3.1 while preserving its voice quality? I’d appreciate any tested workarounds or guidance from the Google team.