Gemini 3.8 Live Avatar gives AI customer service a lip-synced face in 97 languages

Google has put a real-time, lip-synced video face on its speech-to-speech model and sold it first to service desks and showrooms. Every frame carries a SynthID watermark; no launch document puts a price on the face.

Customer service has spent two decades talking to people through a blank box: a phone menu, a chat bubble, a synthetic voice with nothing behind it. On September 25, Google made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise, and the product is a direct attempt to fill that box with a face. The avatar speaks, moves its lips in time with its words, shows expressions and keeps talking while it looks things up in the background. It works in 97 languages and switches between them in the middle of a call when the customer does. Google pitches it at insurance claims, hotel check-ins and car shopping rather than at demo stages, which says a good deal about where the company thinks the money in voice AI sits.

Under the face sits Gemini 3.8 Live, the native speech-to-speech model Google announced the week before, together with an Extended Thinking variant. Speech-to-speech means the model hears audio and answers in audio without first turning the conversation into text and back again. Older voice bots were built as a relay: one system transcribed the caller, a second decided what to say, a third read the answer aloud. Every handoff added delay and stripped out tone, which is why those bots talked over people, missed interruptions and left long, mechanical silences. A native model removes the relay and keeps the rhythm of speech inside a single system.

Live Avatar adds a streaming video layer to that loop. What separates it from video overlay techniques, where a face is animated after the audio already exists, is that audio and video are generated in the same live inference path rather than as sequential steps. That coupling is what keeps latency low enough for a natural back-and-forth, and it is why the lips land on the syllables instead of trailing them. The model also reads live camera feeds and screen shares while it talks, and it can highlight elements on the customer’s screen, so a caller can show a problem as well as describe it and the agent can point back at the answer.

The product did not appear from nowhere. Google first previewed Live Avatar at Google Cloud Next 2026, and the Live line goes back to 2024, when Gemini Live arrived as a voice mode on phones and the Multimodal Live API opened to developers alongside Gemini 2.0 that December. Each version pushed the same idea a little further: fewer turns that feel like turns, fewer pauses that give the machine away. The consumer-facing post introducing the avatar went up on September 24, signed by Shuo-yiin Chang, a research scientist, and CJ Zheng, a software engineer, on behalf of the Gemini Audio Team. The enterprise general availability followed a day later on the Google Cloud blog, which tells you which audience Google expects to pay for it.

For a customer-service deployment, the face matters less than what happens behind it. Gemini 3.8 Live executes tool and API calls asynchronously: it acknowledges the request, keeps the conversation going and fetches an order, checks a policy or queries inventory in the background. Google says it recovers from interruptions without dropping the context or the backend transaction in progress, which is the failure that most often sends a frustrated caller back to the start of a phone tree. The difference is easy to underestimate. A bot that goes quiet for six seconds while a database answers feels broken; one that says it is checking and carries on talking feels like a clerk.

Google’s showcase is an insurance claim. The customer points a phone camera at the damage and talks it through, the agent fills in a claim record as they speak, and a team of background agents built with Google’s Agent Development Kit checks the policy and assembles the packet an adjuster will need. The code for that agent team is linked from the launch material as an open-source example, a sign that Google wants developers to copy the architecture as much as the face. The company’s other example is more ordinary: a hotel guest checking in. Both describe the same design. The avatar is the front end of a workflow, not a talking head reading a script.

Language handling is the second pillar. Detection is automatic, and Google says the avatar can move across its 97 languages inside the same session without degrading video fidelity or introducing visual drift, with lip movement adapting to the new language rather than keeping the mouth shapes of the old one. The switch happens without a handoff, a system reload or a noticeable pause. In markets where customers move between two or three languages inside a single call, which is the daily reality of support lines in India, Nigeria or the Philippines, that is the difference between a usable system and one that forces people to pick a language from a menu before they can explain why they called.

Equal AI, a company building a personal AI assistant in India, gives the clearest picture of scale, although its numbers describe the voice model rather than the avatar. The company says it already handles more than a million live calls a day across nine Indian languages, and its chief executive, Akhilesh Damaraju, credits Gemini 3.8 Live with better interruption handling, more reliable multilingual conversations and steadier tool calls. That volume would be economically impractical with human agents at the same availability and scheduling coverage. The design assumption behind a deployment like that is not proof of concept. It is throughput, and at that volume every second shaved off a conversational turn adds up.

The other names in the launch point at different ends of the market. Cox Automotive has built a conversational avatar for Autotrader that lets shoppers describe the car they want in their own words, highlights matches on screen and calls tools for search, comparison and financing; Marianne Johnson, Cox Automotive’s chief product officer, frames it as a replacement for filters and menus. Salesforce says its AI Research team is working with Google to combine Gemini 3.8 Live with Agentforce, its agent platform for service, and Bob Van Osten, vice president of product for Agentforce, frames the collaboration as a search for new ways to carry a service case from first contact to resolution. Specs, which builds its own AI assistant, points through software engineer Eric Walsh to the updated voice activity detection and lower latency it measured in testing.

Every second of generated audio and video carries a SynthID watermark. The system, introduced by Google DeepMind in 2023 for images and later extended to text, audio and video, embeds a signal that people cannot see or hear but software can detect. Google designs it to survive common transformations such as compression and re-encoding, so a clip exported from a support call and posted elsewhere can still be checked for its origin. It is the most concrete safety commitment in the launch, and it applies to every avatar, preset or custom. The second safeguard is procedural: custom avatars sit behind an enterprise allowlist and a separate verification process, which is Google’s answer to the obvious worry that a single photo and a voice sample are enough to clone a real person.

The watermark also lands at a useful moment for the European side of the business. Live Avatar ships with endpoints in the United States and the European Union, and the EU’s AI Act includes transparency rules, applicable from August 2026, that require people to be told when they are dealing with an AI system and require synthetic audio and video to be marked in a machine-readable way. SynthID covers the second obligation. The first one belongs to whoever deploys the avatar, because a watermark that nobody can see does nothing to tell a caller that the friendly face on the screen is generated. Google can supply the marking; the disclosure has to come from the bank, the insurer or the hotel chain that puts the face in front of its customers.

What the launch withholds is the price. Neither announcement quotes a figure for the avatar layer; the Cloud post points to Google’s general generative AI pricing page and to provisioned throughput, and both provisioned capacity and custom avatars go through a sales representative. Organizations can pick from a curated library of preset faces, or build their own from a single reference photo, an audio sample and a set of system instructions, but that custom route is gated rather than self-service. Extended Thinking for the Live model remains in private preview, which caps its reasoning depth compared with Google’s other Gemini models for now. Latency figures and concurrent session limits are not published. The restricted profile is consistent with how Google stages enterprise releases, with pricing and capacity negotiated rather than posted.

The avatar also arrives in a category that already exists. Companies such as Synthesia, HeyGen and D-ID built businesses on generated presenters, largely for training, marketing and localized video that is rendered in advance rather than live. Google’s pitch is different in one respect: the face is attached to the same model that listens, reasons and calls the tools, inside the console where enterprises already build their agents. It sits next to a wider family of audio models that includes Gemini 3.5 Transcribe, Gemini 3.5 Live Translate and two text-to-speech variants, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. For a company already running Gemini agents, adding a face becomes a configuration choice rather than a separate vendor contract, and that bundling is the competitive threat the specialists will have to answer.

There are open questions the launch material does not address. A face raises expectations a voice does not: callers read expressions, and an avatar that smiles while refusing a claim may land worse than a neutral voice delivering the same answer. Bandwidth matters too, since a streaming video agent asks more of a customer’s connection than a phone line does, and many of the markets where multilingual support is most valuable are also where mobile data is most expensive. None of this is measurable from a launch post. It will show up, or not, in the satisfaction and completion figures of the first large deployments.

Gemini 3.8 Live with Live Avatar is available now in the Gemini Enterprise console and through the Gemini Live API, with the Agent Development Kit as the developer path for streaming agents. Extended Thinking has no published timeline beyond its current private preview group. The enterprises deploying it will answer a question the technology does not: whether removing the visual signal that marks an automated interaction improves the experience for the people on the other end, or removes the last honest indicator that one is happening.

Tags: artificial intelligence, Google, Gemini, speech recognition, customer service, enterprise AI