Google Adds Talking Live Avatars to Gemini 3.8 Live for Enterprises

Google just gave its newest live conversation model a face. Gemini 3.8 Live with Live Avatar became generally available on September 24 for Gemini Enterprise customers, one week after the audio-focused Gemini 3.8 Live models launched and months after a first look at Google Cloud Next 2026. The update generates a 24-frame-per-second video of an animated person or cartoon character whose mouth and expressions stay locked to the spoken words, pairing the existing speech model with low-latency streaming video in a single pass.

Live Avatar listens to any speech in real time, watches live camera feeds or screen sharing at one frame per second, and responds swiftly with a 24 kHz voice and a real-time moving face. Lip movements remain in sync even when moving from English to Japanese or any of 97 languages, and Google claims the quality never decreases or loses crispness. It can also display more information on the screen as the avatar continues to speak. Furthermore, affective dialog is the default, so the emotions will alter subtly in tandem with the tone of the words you’re saying.

Businesses can choose from a library of pre-made faces, ranging from cartoon characters to very lifelike humans, and have them paired with a matching voice. Any organization on the approved list may simply provide a still photo, and the system will generate a personalized talking version that looks just like a live person while maintaining the original branding. In one instance, the avatar checked in a hotel guest in the middle of a conversation as the system ran booking tools in the background without halting the chat. Before you realize it, Live Avatar is being used in real life by shopping assistants and multilingual support representatives.

Those fluid interruptions function exactly the same as in the audio-only version, so a caller can simply jump in mid-sentence and the context remains intact, including any background work the avatar has already begun. There is no separate transcription step between what you say and what the face replies – this is all native speech-to-speech stuff.

Every second of music and video has an invisible watermark known as a SynthID, which later programs can use to verify that the clip was created by Live Avatar. Google also added some more identification precautions and keeps custom faces under review. Sessions are only a few minutes long (rather than hours as we would prefer), and they are slightly more expensive than using voice alone. As with all of those large models, the occasional delusion or slowness is still possible.

Gemini 3.8 Live with Live Avatar is now live on endpoints in the United States and the European Union, complete with the necessary configuration and typical enterprise security. The more powerful Gemini 3.8 Live Extended Thinking version is still in private preview and does not yet have the face feature. For the time being, the talking heads are entirely internal in company accounts, intended at support desks, product demos, virtual receptionists, interactive assistance that get to see what a customer is pointing to while they speak, and so on.