Gemini Live's Custom Voices Have Quietly Degraded Since the 3.1 Flash Update
Google's Gemini 3.1 Flash Live update delivered faster response latency and a larger conversational context window — but it came with an unacknowledged side effect. Users have begun reporting significant degradation in Gemini Live's custom voice options since the update rolled out, with complaints ranging from slowed speech cadence and shifted accents to audio artifacts and wildly inconsistent voice previews.
9to5Google's own testing confirmed the regression. The "Capella" British accent voice — previously one of the most distinct and consistent options — changed dramatically after the update, bearing little resemblance to its earlier behavior. The pattern points to the model upgrade inadvertently destabilizing voice tuning behavior that had been reliable in prior versions. Google has not publicly acknowledged the issue.
For developers and users who've shipped or rely on voice-first Gemini integrations, this is a real-world reminder that model upgrades can quietly break downstream UX — even when the headline metrics improve. Voice consistency is still an unsolved challenge in production real-time AI, and this regression is worth tracking closely before launching voice-dependent products on the current Gemini Live stack.