"Gemini 3.1 Flash Live only manages 36.1 percent in this test, while audio models not designed for conversation can exceed 50 percent in the MultiChallenge."
"The outputs from this model will have SynthID watermarks, which are not perceptible to human listeners but can be detected if someone tries to pass off Gemini AI speech as real."
"Google has partnered with companies like Home Depot and Verizon to test the model, receiving glowing reports on how well 3.1 Flash Live can mimic human speech."
"Developers can now access the model in AI Studio, the Gemini API, and Gemini Enterprise for Customer Experience, with the new conversational AI rolling out in those products."
Gemini 3.1 Flash Live shows improved performance in handling audio interruptions and hesitations, achieving 36.1 percent in the Audio MultiChallenge. Despite lower scores compared to non-conversational models, it aims to sound more human-like. Google has integrated SynthID watermarks to ensure authenticity in AI-generated speech. Partnerships with companies like Home Depot and Verizon have yielded positive feedback on its realistic speech capabilities. The model is now accessible through AI Studio, Gemini API, and Gemini Enterprise, with rollout in Gemini Live and Search Live features.
Read at Ars Technica
Unable to calculate read time
Collection
[
|
...
]