Gemini 3.8 Live now generates a video avatar that lip-syncs across 97 languages: Reveneau

Why it mattersA voice agent can now hand off a slow query without going silent, and the same avatar can switch language mid-sentence without breaking eye contact, which changes what an enterprise support flow can look like.

A voice agent has always had one honest problem: it goes silent the moment it needs to look something up. Google DeepMind's Gemini 3.8 Live with Live Avatar, released today, closes that gap with two changes.

The DeepMind post says the model pairs live dialogue with real-time video generation. Users see a face that listens, looks, and speaks, with precise lip-sync and natural expressions. It is a follow-on to last week's Gemini 3.8 Live launch and ships today inside Gemini Enterprise.

Background tool calls with continuous presence

The first change is asynchronous tool execution. DeepMind's own words: Live Avatar can trigger tool calls and fetch data in the background while continuing active dialogue, handling complex tasks while ensuring an uninterrupted conversational flow. The example in the post is a guest hotel check-in that continues talking while the agent looks up a booking record.

Any voice agent built on a slower back-end query has had the same behaviour for years: a pause while the query runs. Removing that pause means the caller keeps hearing a voice for the whole call, not a hold silence in the middle.

One avatar, 97 languages, one lip-sync engine

The second change is multilingual speech-to-speech that keeps the video stable. DeepMind says the avatar switches between 97 languages mid-conversation, with lip-sync and expressions adapting without introducing visual drift. The demo shows a single avatar swap between languages inside one turn.

Every enterprise that supported multilingual voice previously either ran a separate pipeline per language, or lost lip-sync the moment a caller switched languages mid-sentence. A single avatar that carries expressions across the switch changes what a globally deployed agent can look like.

What the release actually opens

Live Avatar ships in Gemini Enterprise with a library of preset avatars. Organisations can also generate a custom avatar from a single high-quality reference image, which DeepMind says preserves reference likeness, brand styling, or character identity. That custom path is currently gated behind an enterprise allowlist.

Every output, audio and video, is watermarked with SynthID. DeepMind says the watermark is imperceptible and woven into both streams so AI-generated content stays detectable. The post links to the model card for the safety details.

For a team already building an agent on Gemini Live, this is a direct swap. For a team using another provider for voice, the avatar plus the async tool loop is the pair to weigh, since one without the other still leaves the caller staring at a static frame or hearing a stall.

Source

Introducing Gemini 3.8 Live with Live Avatar, Google DeepMind, 2026-09-24

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.