Google says its new realtime voice model improves noisy-environment understanding, long conversations and function calling, and it's rolling into Gemini Live, Search Live and AI Studio. Voice creators can test it for lower-latency spoken interactions.
Google is pitching this release as a practical upgrade for spoken interfaces, not just a benchmark refresh. The launch materials in the main announcement show Gemini 3.1 Flash Live hitting 90.8% on ComplexFuncBench audio and 95.9% on Big Bench Audio speech reasoning, with the included charts
and
positioning it ahead of earlier Gemini native-audio versions.
For creators building voice-led experiences, the more useful change is behavioral. Google's product thread says the model keeps track of longer conversations, understands details in noisy settings, and makes function calling more reliable, which maps directly to voice agents that need to listen, remember context, and trigger tools without awkward retries. The availability update says it is already rolling into Gemini Live and Search Live, while developers can start testing it in AI Studio and dig into the fuller product write-up via Google's overview.
Introducing Gemini 3.1 Flash Live, our new realtime model to build voice and vision agents!! We have spent more than a year improving the model + infra + experience, the results? A step function improvement in quality, reliability, and latency.
Say hello to Gemini 3.1 Flash Live. 🗣️ Our latest audio model delivers more natural conversations with improved function calling – making it more useful and informed. Here’s what’s new 🧵 Show more
