Google's new Live Avatar gives Gemini a face and 97 languages

Talking to a text box or a bouncing audio orb might soon feel like a relic of early AI.

Following its rollout of Gemini 3.8 Live, Google just introduced Gemini 3.8 Live with Live Avatar. The feature pairs near real-time video generation with native speech models to create interactive visual personas that listen, look, and speak back to you.

It is officially rolling out to Gemini Enterprise users starting today.

Multitasking avatars across 97 languages

Instead of the stiff puppets we usually see in virtual customer service, Live Avatar focuses on precise lip-syncing, fluid turn-taking, and expressive faces.

Because conversation is naturally multimodal, the system processes both audio and visual inputs at the same time, adjusting what it says and how it looks on the fly.

The global reach here is pretty striking. It supports native speech-to-speech synchronization across 97 languages. Google claims you can switch languages mid-conversation, and the avatar will adapt its lip-sync and facial expressions instantly without visual drift or drops in fidelity.

For enterprise tasks, the biggest technical upgrade is asynchronous tool execution. Typically, when an AI agent queries a database, the conversation grinds to a halt. Live Avatar can fetch data or trigger API calls in the background while keeping up active, face-to-face small talk.

In one demo, Google showed an avatar handling a hotel check-in: pulling up guest records behind the scenes while chatting continuously with the user without any awkward pauses.

Brands can either pick from a library of preset characters or generate an animated persona from a single high-quality reference photo. Custom generation is currently limited to an enterprise allowlist, and all audio and video outputs are embedded with imperceptible SynthID watermarks.

Why it matters

For enterprise workflows like customer support, onboarding, and virtual retail, this is a massive upgrade over sterile robotic voices. Being able to spin up a branded avatar from just one photo lowers the barrier for companies that could never afford custom 3D animations.

Across the industry, Google is shifting the frontier from mere multimodal understanding to real-time multimodal embodiment. Rivals will now have to prove their live voice assistants can also look human without ballooning latency.

We think the real breakthrough here is asynchronous tool execution. While many avatar projects obsess purely on photorealism, keeping a natural conversation going while querying back-end systems is what actually makes an agent usable in real-world business.

That said, the compute cost of running continuous near real-time video streaming will be a major barrier to widespread adoption. Furthermore, even with SynthID watermarking, being able to clone a responsive avatar from a single static image will inevitably keep identity and security watchdogs on high alert.