Google has announced two new models to put an end to the dilemma that has long defined voice AI assistants: “speed first, or depth first?”: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The company said it has built an architecture that performs multi-step reasoning in the background during a live voice conversation without making the user wait.
Two models, one conversation flow
The new architecture consists of two specialized models. Gemini 3.8 Live is optimized for very low-latency direct conversations; it will be used in scenarios that require fast responses, such as language practice, instant translation, call routing, and voice search. Gemini 3.8 Live Extended Thinking, meanwhile, was developed for tasks such as complex planning, high-level problem solving, and code analysis; it can process reasoning steps in the background while continuing to talk with the user.
Developers will be able to define hybrid switching between the two models via Google AI Studio and Vertex AI APIs. While the standard Live engine responds to a simple question, for requests such as financial statement analysis or end-to-end travel route optimization, the system automatically switches to the Extended Thinking layer.
Background reasoning that ends the silence
Google engineers say they solved this transition with a technique they call “Concurrent Cognitive Decoupling.” While the model runs a Python compiler in the background, cross-checks five separate sources, or examines a shared video recording frame by frame, the dialogue doesn’t break. Robotic “please wait” announcements are replaced by contextual interjections such as “I’m checking the tax rates in the table; meanwhile, would you like me to factor in last month’s spending as well?”
Thanks to asynchronous tool calling, the user can interrupt before a function responds and add a new parameter. When the user says, “no, I meant the 2026 budget, not 2025,” the calculation engine running in the background abandons the old branch and switches to the new route.
Agents that see the screen and hear the voice
The architecture is not limited to voice; the local multimodal infrastructure also processes video, screen sharing, and images in real time. For example, when a software engineer sharing their screen live says, “I can’t find the memory leak here,” the model can evaluate both the question mark in their tone of voice and the profile data on the screen, then point to the line in the call graph where the pointer is not released.
Google says it offers the models through Google Workspace, the Gemini mobile app, and Google Cloud. Among the use cases the company highlights are voice agents that query CRM and logistics databases in call centers without putting the customer on hold, doctors doing literature reviews without using their hands, and education assistants that watch the problem a student is solving through the camera and give hints.
Google is positioning this dual architecture as the new standard in voice AI. How consistently the models perform in the field will be shown by the experience of developers and enterprise customers.
Source: WebTuring

Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.