Gemini's live voice model gets a face

Google has released Gemini 3.8 Live with Live Avatar, which adds a near real-time video persona to its native live dialogue models. The feature pairs streamed video generation with speech, so an agent listens, looks and answers with a face that moves in step with what it says. It follows the launch of Gemini 3.8 Live the week before and is available in Gemini Enterprise.

The avatar is not a separate layer bolted on after the voice. Google describes the video as natively coupled to the dialogue model, which takes in visual and audio input at the same time and responds with expressive audio and video. Tool calls run asynchronously: the avatar can fetch data in the background while the conversation carries on. Google's own demonstration is a hotel check-in.

Language coverage is the claim most likely to matter to companies running customer service across markets. According to Google, lip-sync and facial expressions adapt when a speaker switches language mid-conversation, across 97 languages, without degrading video fidelity or introducing visual drift.

Brand control comes in two tiers. There is a library of preset avatars, and organisations can also generate a fully animated custom avatar from a high-quality reference image, preserving a likeness, a brand style or a character's identity. That second route is restricted for now: custom avatar creation is available only through enterprise allowlisting.

The restriction sits next to the safeguard Google puts first. All output from its AI products is watermarked with SynthID, an imperceptible mark woven into both the audio and the video and intended to keep generated content detectable and limit misattribution. The gate on custom avatars and the watermark on all output are the two controls Google names for a system that can animate a likeness from a single image.