Gemini Live's voices don't sound like they should

Google's Gemini Live voice assistant is experiencing widespread quality degradation following the recent Gemini 3.1 Flash Live update, with users reporting that voice previews bear little resemblance to actual conversation experiences.

What's happening

According to 9to5Google's testing, multiple voice options are exhibiting noticeable changes in cadence, tone, and accent consistency. The "Capella" voice option, which mimics a female British accent, has degraded the most visibly, though other regional voice presets show similar deterioration patterns.

Key issues include:

• Speech pattern changes: Voices sound significantly slower in preview than during actual use
• Accent drift: Preset accents shift or hybrid between different regional options during conversations
• Tone alterations: High-pitched voices are being toned down inconsistently
• Audio artifacts: Users report prominent crackles, pops, and hisses in voice output
• Inconsistent resets: Accents briefly stabilize after app resets, then slowly morph into hybrid versions

Scope and impact

The issue appears sporadic and reproducible primarily with Gemini Live's dedicated voice feature. Voice controls and Android Auto integration remain largely unaffected, suggesting the problem is isolated to the Live conversation system rather than broader text-to-speech infrastructure.

Users have documented these issues on Google's Support Forums, with reports indicating the problem has persisted for months across multiple voice options. The timing correlates with the Gemini 3.1 Flash Live rollout, though Google has not confirmed whether model updates are the root cause.

Google's response

9to5Google has reached out to Google for comment but has not yet received a response. Google has not publicly acknowledged the voice quality issues on its official channels.

What this means

This appears to be a quality assurance gap in Google's voice synthesis pipeline, possibly stemming from model updates that alter TTS outputs without corresponding quality validation. For users relying on Gemini Live for consistent voice interactions—particularly those who selected specific accents—the experience has become unreliable. The gap between preview audio and actual performance suggests either voice model drift during inference or insufficient testing before deployment of the 3.1 Flash update.