Google’s Live Avatar keeps talking while it fetches data

Enterprise agents built on Gemini 3.8 Live can now talk to customers through an animated face. Google made Live Avatar generally available in Gemini Enterprise on September 24. In the company’s hotel check-in demo, the avatar keeps talking to the guest while tool calls fetch the check-in details in the background.

Google’s Gemini Audio Team built Live Avatar as an extension of Gemini 3.8 Live, the conversational model that shipped the week before. It couples the model’s dialogue with low-latency video, so one agent listens, sees and speaks through a persona on screen. Testingcatalog reports the target uses as customer service and guided walkthroughs. Google’s demos showed several avatar characters, a near-real-time conversation, the hotel check-in and a switch between languages in the middle of one conversation.

Google’s Cloud documentation describes how the pieces fit together. The model runs over bidirectional WebSockets and takes in speech, live camera feeds and screen broadcasts with sub-second latency. It returns 24 kHz audio and a synchronized 24 FPS avatar. A developer asks for the video with response_modalities=["VIDEO"] and receives an MP4 stream generated directly from the model’s output, which is why the lip movements and expressions line up with the synthesized speech as it happens. The previous model, gemini-live-2.5-flash-native-audio, lists Live Avatar synthesis as “Unsupported” and outputs only audio and text.

The hotel demo depends on tool calling. In 3.8 Live, function calls can be asynchronous and non-blocking, or blocking with auto-cancel (behavior="BLOCKING"), and audio keeps playing while the backend handles the request. In practice the avatar stays in the conversation while a reservation lookup runs, without going quiet. The 2.5 model offered asynchronous and synchronous calling.

Turn-taking has changed too. When the user starts speaking, 3.8 Live applies a “polite interruption downgrade” and waits. The 2.5 model interrupted immediately and could talk over the user. Affective dialogue and proactive audio are on by default in 3.8 Live, while 2.5 needed explicit configuration for both. For camera and screen input, a configurable media_resolution (LOW, MEDIUM, HIGH) controls how many tokens the visual stream uses, where 2.5 had a fixed per-frame budget. Developers can also bias transcription toward their domain with custom_vocabulary.

Google says the native multilingual speech-to-speech sync covers 97 languages. Lips and expressions adapt when the spoken language changes, and the system is built to switch languages without losing video fidelity or showing visible drift.

The underlying model arrived on September 15, when Google’s developer blog put Gemini 3.8 Live and 3.8 Live Extended Thinking into the Gemini API and Google AI Studio. Google said the Extended Thinking variant ranked first on Artificial Analysis’ Speech-to-Speech leaderboard. Google pitched the models as a simpler alternative to cascaded pipelines, which chain separate speech recognition, language and speech synthesis models. The post also listed alphanumeric precision for things like confirmation codes and claim numbers, and named integration partners including Agora, LiveKit, LangChain, Pipecat and Vercel. Prices were $0.005 per minute of audio input and $0.018 per minute of audio output, about 30 cents and $1.08 per hour of conversation. Those rates were published for audio before the avatar launched.

Companies can pick from a library of preset avatars or build a branded character from a high-quality reference image. Google says the result can keep the source likeness, visual style or character identity while becoming fully animated and responsive. Custom avatars are more restricted than the core product and currently require enterprise allowlisting. The release includes safeguards meant to respect identity, and SynthID watermarks both the audio and the video output. Google says the imperceptible marker is there to keep AI-generated material detectable and to reduce misinformation or misattribution.