Google Adds a Real-Time Speaking Avatar to Gemini 3.8 Live in 97 Languages

Google DeepMind on 24 September added a visual persona to its live dialogue system. Gemini 3.8 Live with Live Avatar generates video in near real time alongside speech, so the system listens, sees, and answers as an animated face with lip movement matched to the audio it produces.

The engineering claim worth attention is synchronization under language switching. Google says the avatar lip-syncs and handles turn-taking across 97 languages and detects a change in spoken language without losing alignment between the audio and the video it is rendering, as The Tech Portal reported. Two hard technical problems underlie that claim: keeping generated video coherent at conversational speed, and keeping a speech-to-speech model stable when the phoneme distribution (the mix of speech sounds it must produce) shifts mid-conversation.

The second feature matters more for anyone building with it. The system supports asynchronous tool calling during an active conversation, which means the avatar can query a backend while continuing to speak instead of going silent. That is the difference between a demonstration and a usable support or sales interface. Enterprises on an allowlist can generate a custom avatar from a single reference image.

Google has not published latency figures, and no independent evaluation has tested the avatar under poor network conditions, interrupted speech, or heavy tool-call load. The available details come from Google's announcement and its Gemini 3.8 Audio model card. One safety decision is concrete and verifiable by outside parties: every audio and video output carries SynthID watermarking, Google's imperceptible provenance signal. The model is available now through Gemini Enterprise rather than the consumer app, which indicates who Google expects to pay for it.