Live Avatar is an animated on-screen character that speaks a chatbot’s replies aloud while lip-syncing and shifting expression in real time. The Verge reports it shipped with a Gemini 3.8 Live update and is limited to Gemini Enterprise customers.
An animated face on a chatbot
A conversational avatar is a synthetic face, rendered on screen, that moves in time with a spoken response generated by an AI model. Instead of reading text or listening to a disembodied voice, the user watches a character whose mouth forms the shapes of the words being spoken and whose eyes, brows and head move while it talks and while it listens.
According to The Verge, Google’s version of this is called Live Avatar, and it is part of an update to the Gemini model’s live conversational mode. The publication reports that the avatar lip-syncs and displays different facial expressions during a conversation, and that access is restricted for now to business customers on Gemini Enterprise rather than the general consumer app. The Verge’s account also refers to a figure of 97 in connection with transitions the avatar can make, but the passage available here is cut off and what that number counts is not stated.
Nothing about the underlying language model changes because a face is attached to it. The avatar is a presentation layer: the same reasoning, the same errors, the same limits, wrapped in something that looks like a speaker.
Roots in digital humans and voice assistants
Two separate lines of work meet in a product like this. The first is computer-animated speech. Since the 1990s, animation and games research has worked on mapping phonemes — the individual sounds of speech — onto visemes, the corresponding mouth shapes, so that a rigged character can appear to pronounce arbitrary text. Film and game studios industrialised this, first with hand-keyed animation and later with automated systems that drive a facial rig directly from an audio track. Broadcast and customer-service “digital humans” followed, usually running from a script or a pre-recorded performance.
The second line is real-time spoken dialogue. Voice assistants of the 2010s worked in stages: record the user, transcribe to text, run the text through a system, synthesise a reply. The delay between each stage made anything resembling natural turn-taking difficult. More recent multimodal models process audio, and in some cases video, as a continuous stream and emit audio in return, which narrows the gap between a person finishing a sentence and the system beginning to answer.
Once a system can hold a low-latency spoken exchange, adding a face becomes a rendering problem rather than a scripting problem. The animation can be driven by the audio the model is producing at that moment. Several firms have offered avatar layers of this kind to businesses for some years, typically for support desks, training material and kiosks.
Current mechanics
In general terms, a live avatar system needs three things working together within a few hundred milliseconds. It needs the spoken reply: audio synthesised from the model’s output, or generated directly by a speech-capable model. It needs facial motion derived from that audio, so that mouth shapes align with the sounds and the movement does not drift out of sync as the sentence runs on. And it needs a rendering pipeline that produces video frames and delivers them to the user’s screen alongside the audio.
There are two broad approaches to the middle step. One drives a three-dimensional character rig, adjusting a set of named controls — jaw, lip corners, brow height, eyelids — that an animator would recognise. The other generates video frames directly with a neural model conditioned on audio and a reference image. The first tends to be more predictable and cheaper to run; the second can look more photographic but is harder to constrain. Which approach Google uses for Live Avatar is not stated in the material available here.
Expressions add a further layer. Beyond the mouth movements required by speech, a system may select among a library of expressive states and blend between them, which is the kind of behaviour The Verge describes when it refers to the avatar transitioning between states during a conversation. How such states are chosen — from the sentiment of the text, from explicit tags in the model’s output, or by some other means — varies between systems and has not been detailed for this one.
Enterprise-first release is a common pattern for features like this. Business deployments are smaller, contractually bounded, and easier to monitor than a consumer launch, and the compute cost of streaming rendered video to every user is considerable.
Common misunderstandings
The most persistent error is treating the face as evidence of understanding. An expressive avatar that appears to react with interest or concern is performing an animation selected by software; there is no inner state behind it. Decades of research on computer-generated characters point the same way: people attribute intention, competence and even honesty to a system that behaves socially, and that attribution is not calibrated to whether the system is right.
A second misunderstanding is that the avatar makes the model more capable. It does not. Whatever the model gets wrong in text, it will get wrong through a face, and a confident delivery can make an incorrect answer harder to question.
Third, people often assume a rendered face means a real person was recorded. Synthetic characters need not resemble anyone; some are deliberately stylised. Conversely, systems that do reproduce a real person’s likeness or voice raise separate consent and rights questions, which is why the distinction matters.
Finally, more realism is not automatically better. The “uncanny valley” — an idea from 1970s robotics research — describes the drop in comfort that occurs when a synthetic figure comes close to human appearance without quite reaching it. Small mismatches between audio and lip movement, or a blink at the wrong moment, can be more unsettling than an obviously cartoonish design.
Further ground to cover
Readers who want to follow this properly should look at three things. Google’s own product documentation for Gemini Enterprise is the only authoritative account of what the feature does, which customers can use it and what controls administrators have. Academic literature on embodied conversational agents and speech-driven facial animation, published through venues such as ACM and IEEE, covers both the techniques and the measured effects on user trust. And regulatory material — notably the transparency provisions of the EU AI Act, which require that people be informed when they are interacting with an AI system — sets out the disclosure duties that apply as synthetic faces spread into customer-facing services.
Frequently asked questions
What is Google’s Live Avatar?
Live Avatar is an animated character that appears during a live spoken conversation with Google’s Gemini model and moves as it speaks. The Verge reports that it lip-syncs to the generated speech and shows a range of facial expressions while the exchange takes place. It was introduced with a Gemini 3.8 Live update. It is a visual layer over the model’s spoken output rather than a new capability of the model itself.
Who can use Live Avatar?
According to The Verge, the feature is currently available only to Gemini Enterprise customers, meaning organisations on Google’s business tier rather than users of the free or consumer Gemini app. Whether it will reach a wider audience, and on what timescale or terms, is not stated in the reporting available. Enterprise-first releases are common for features that are costly to run or that benefit from controlled deployment.
Does an AI avatar have real emotions?
No. The expressions an avatar displays are animations chosen and blended by software, usually driven by the content of the response or by the audio being produced. There is no feeling behind them and no internal state that the face reports on. Research on social interfaces consistently finds that people read sincerity and competence into such behaviour anyway, which is precisely why the distinction is worth stating plainly.
How does lip-syncing work in AI avatars?
Broadly, the system maps the sounds in the synthesised speech onto corresponding mouth shapes, known as visemes, and drives a facial model to form them in sequence. Some systems animate a three-dimensional character rig; others generate video frames directly with a neural network conditioned on the audio. The whole chain must run fast enough that mouth movement and sound stay aligned during live conversation. The method used for Live Avatar has not been detailed.
Are companies required to disclose that a face is AI-generated?
Rules vary by jurisdiction. In the European Union, the AI Act contains transparency obligations requiring that people be told when they are interacting with an AI system, and further duties around artificially generated or manipulated content. Other countries have narrower rules, often focused on likeness rights or advertising standards. Specific obligations depend on the deployment and the territory, and are best checked against the current legal text.
Sources and further reading
- The Verge — technology news reporting that carried the account of the Gemini 3.8 Live update and the Live Avatar feature.
- Google’s Gemini Enterprise product documentation — the authoritative reference for feature availability, administrative controls and supported use.
- ACM and IEEE digital libraries — peer-reviewed research on speech-driven facial animation, embodied conversational agents and user trust in synthetic characters.
- The European Union’s AI Act — statutory text setting out transparency obligations for AI interaction and synthetic content.
Surfaced from the rss:verge signal “an AI product launch”. AI-assisted draft, editorially reviewed.
