Google launches Gemini 3.1 Flash Live real-time voice AI
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Google announced Gemini 3.1 Flash Live, a new AI audio model built for real-time conversation, rolling out in Gemini Live and Search Live (a feature of AI Mode) starting today, with developer access via AI Studio and the Gemini API.
- Gemini 3.1 Flash Live targets more natural cadence and reduced latency, though Google has not disclosed a specific delay despite researchers identifying 300 milliseconds as the optimal limit for speech perception.
- The model topped benchmark tests including ComplexFuncBench Audio, Big Bench Audio (a 1,000-question reasoning set), and Scale AI's Audio MultiChallenge — where it scored 36.1%, still below the 50%+ that non-conversational audio models achieve on interruption handling.
- Google added SynthID watermarks to all outputs, which are imperceptible to human listeners but can be detected to verify whether a clip was generated by the Gemini model.
- Home Depot, Verizon, and other unnamed partners piloted the model and gave glowing reviews in Google's announcement, signaling imminent deployment in customer-facing phone interactions.
- Gemini Enterprise for Customer Experience is positioned as the agentic shopping toolkit layer, giving businesses a direct path to deploy the model in retail and support contexts.
Why it matters: With Home Depot and Verizon already piloting Gemini 3.1 Flash Live for customer-facing calls, the technology is moving from demo to production voice channels — but the 36.1% MultiChallenge score (versus 50%+ for non-conversational audio models) shows real-time AI voice still struggles with interruptions, meaning callers will likely hear increasingly human-like bots that nonetheless glitch when conversations get messy.
