Google Gives Gemini a Talking, Lip-Synced Avatar

Google has given its live AI assistant a face. On 24 September the company made Gemini 3.8 Live with Live Avatar generally available in Gemini Enterprise, adding a generated, lip-synced video presenter to the speech its live models already produce.

The pitch is a single conversation that looks and sounds like a person. The avatar, an animated human or a cartoon character depending on what a business chooses, speaks and moves its mouth in step with the words, rendered at 24 frames a second across 97 languages. The speech and the video come out of the same pass rather than being stitched together afterwards, which is what keeps the lips matched to the sound.

Under the surface it is the interaction model that matters more than the picture. Gemini 3.8 Live runs native speech-to-speech, so it recovers from interruptions without losing the thread of the conversation or dropping a transaction it was in the middle of. It can also fire off tool calls and API requests in the background while it keeps talking, which is the part enterprises will care about when they wire it into a helpdesk or a booking system.

Google is selling it strictly as an enterprise product for now. It runs only in Gemini Enterprise, on endpoints in the United States and the European Union, with provisioned throughput and the compliance and data-governance controls that large customers demand. The video costs an extra dollar per million output tokens on top of the audio a customer already pays for. Every stream carries an imperceptible SynthID watermark, Google's tag for marking AI-generated media so it stays detectable later.

The launch rounds out a busy month for Google's voice and video work. In September the company also shipped new text-to-speech models that can clone a voice from about thirty seconds of audio, and it confirmed that Gemini 4 has entered post-training. A talking avatar is the customer-facing end of the same push.

Whether people want to talk to a synthetic face is a separate question from whether the technology works. The watermark suggests Google knows the answer is not obviously yes. A lifelike presenter that never blinks wrong is impressive engineering and, for some callers, a small uncanny jolt they did not ask for.