Google's real-time avatar layer for Gemini 3.8 Live — a talking on-screen persona that listens, sees, and answers with lip-synced audio and video, across 97 languages, inside Gemini Enterprise.
Last verified · 2026-09-24 · by Moe Ameen
Gemini 3.8 Live with Live Avatar, announced September 24, 2026, is the visual layer on top of Google's real-time voice model. Gemini 3.8 Live could already listen, see, and talk; Live Avatar adds a face to it by pairing near-real-time video generation with speech. The result is a dynamic on-screen persona that responds with precise lip-syncing, natural facial expressions, and fluid conversational turn-taking, processing your visual and audio input at the same time and answering with expressive audio and video together.
The engineering that matters for how it feels: it supports asynchronous tool calling, so it can start a background lookup or API call and keep talking while the data comes back, and it works across 97 languages, adapting the avatar's lip-sync as it switches. All audio and video output carries an imperceptible SynthID watermark so it stays detectable as AI-generated.
On avatars, Google offers a library of preset faces to every enterprise user; creating a custom avatar from a reference image requires enterprise-allowlisted access. The important framing for a creator is where it lives and what it's for. Live Avatar launched inside Gemini Enterprise — Google announced no consumer tier and no price — and it's built for live business use cases: support agents, sales assistants, training, onboarding. A talking face on an ongoing conversation is the product.
That also defines the limit. Live Avatar is a conversational interface, not a content-production tool. It runs a real-time, on-screen exchange, but it exports no captioned short, no carousel, no blog, no newsletter, and no scheduled post. It gives an AI a face for a conversation; it does not give you a library of finished, on-brand video to publish.
The cleanest way to place Live Avatar is by what kind of face it is. Live Avatar renders a face for one live conversation — reactive, ephemeral, one-to-one, and (today) locked to Gemini Enterprise. [Kompozy](/) renders a face for a content calendar — produced in advance, reusable, one-to-many, and available to any creator without an allowlist. Those aren't the same product, and the difference is exactly the seam most creators need closed. What you actually want to publish isn't a chat that happened; it's a recurring on-camera identity your audience recognizes across weeks of posts.
That's the AI Influencer persona pool in Kompozy: you define one or more personas (one primary, many variety faces), and every render is held to that identity. From a single script or source, Kompozy generates [Persona Shorts](/glossary/persona-shorts) — HeyGen talking-head video with a face-locked presenter, native voice, and auto-captions — plus [Persona Frames](/glossary/output-buckets) that composite that avatar inside a brand-exact [HyperFrames](/glossary/hyperframes) template, and it doesn't stop at video: the same idea fans into carousels, quote graphics, photo posts, a blog, and a newsletter, all governed by one [Persona Brief](/glossary/persona-brief) and banned-word filters. Because it's batch and asynchronous rather than live, you produce a week of on-brand content at once instead of talking through a single interaction — then [Autopilot](/glossary/autopilot) schedules and publishes the set across the eight social platforms plus blog and email behind a per-post review gate. Live Avatar is the face that talks back; Kompozy is the face that ships.
It is a real-time avatar layer for Google's Gemini 3.8 Live voice model, announced September 24, 2026. It pairs near-real-time video generation with speech to create an on-screen persona that listens, sees, and answers with precise lip-syncing, natural expressions, and fluid turn-taking, across 97 languages, with SynthID watermarking on the audio and video.
Not really — it is a live conversational interface, not a content generator. It holds a spoken, on-screen exchange but exports no captioned short, carousel, or scheduled post. For a recurring, face-locked presenter that produces publishable video and posts, use a generation engine like Kompozy, which makes Persona Shorts and Persona Frames avatar video and publishes across platforms.
At launch it is available only in Gemini Enterprise, with no announced consumer tier or price. Preset avatar faces are available to all enterprise users; creating a custom avatar from a reference image requires enterprise-allowlisted access. Most individual creators and small brands cannot access it today.
Live Avatar is a face for a live, reactive conversation; a Kompozy persona is a face for produced, publishable content. Kompozy generates avatar video in advance from a script, holds it to a consistent Persona Brief and face-locked identity, fans the same idea into carousels, blogs, and newsletters, and schedules it across nine platforms — none of which a real-time conversational avatar does.