Direct answer: Gemini 3.8 Live with Live Avatar is Google’s new enterprise system for real-time, face-to-face AI agents. It can listen, see, speak and generate a synchronized video avatar while the conversation is happening. It is generally available through Gemini Enterprise, using the gemini-3.8-live model; custom avatars require enterprise allowlisting.[1][2][3]
Fastest way to try it: Open the Gemini Enterprise multimodal live studio, select Gemini 3.8 Live, start with a prebuilt avatar, test a short conversation, and review the privacy and disclosure checklist below before connecting customer data or tools.
Gemini 3.8 Live Avatar at a glance
| Question | Answer |
|---|---|
| Is it available now? | Yes, in Gemini Enterprise |
| Model ID | gemini-3.8-live |
| What can it receive? | Streaming audio, video/camera frames and text |
| What can it return? | Spoken audio, text and synchronized 24 FPS avatar video |
| Languages | Google says it can switch across 97 languages |
| Can it call tools? | Yes, including non-blocking background tool calls |
| Can anyone create a custom avatar? | No. Custom avatar creation is allowlist-only |
| Is generated media marked? | Google says audio and video output carries SynthID watermarking |
Google positions Live Avatar for customer service, interactive walkthroughs, kiosks, virtual tutors and other experiences where a visible agent is more useful than a voice-only assistant.[1][2][3]
What is Gemini 3.8 Live with Live Avatar?
Gemini 3.8 Live is a bidirectional streaming model built for continuous conversations. The Live Avatar feature adds a generated visual persona whose lip movements and facial expressions stay synchronized with the model’s speech. Google’s developer guide lists 24 kHz spoken audio and 24 FPS MP4 avatar output.[3]
The model can process live camera frames or screen sharing while listening to the user. It also supports asynchronous function calling, which means an agent can keep talking while a backend task—such as checking an order or retrieving an account record—runs in the background.[1][3]
This is different from text-to-speech. A TTS model reads a prepared script; Live Avatar handles an open, two-way conversation with visual context and tools. If you only need narration, use the simpler Gemini 3.8 Flash TTS workflow instead.
What you need before setup
Use this preflight checklist:
- A Google Cloud project with billing and the required Gemini Enterprise access.
- Permission to use Gemini Enterprise Agent Platform and its studio.
- Application Default Credentials if you will use the SDK.
- A clear use case, such as concierge help, product guidance or customer intake.
- A list of backend tools and exactly what each tool may read or change.
- A privacy notice and a visible disclosure that the person is interacting with AI.
- Human escalation for complaints, payments, safety issues and uncertain answers.
- Written permission for every face, likeness and voice used in a custom avatar.
Do not start with production customer data. First test with invented records in a restricted development project.
How to try Gemini 3.8 Live Avatar
1. Open the multimodal live studio
Go to the Gemini Enterprise multimodal live studio. If your account cannot open it, verify the active Cloud project, billing, permissions and supported endpoint with your administrator. Google lists US and EU endpoints for the generally available enterprise release.[2]
2. Start with a prebuilt avatar
Use a curated stock avatar for the first prototype. Google restricts custom avatar creation to an enterprise allowlist and verification process, so a reference image alone does not guarantee access.[1][2]
Choose a neutral persona appropriate for the task. Avoid a face or voice that could imply a real employee, celebrity or public official is speaking.
3. Write a narrow system instruction
Copy and adapt this starter instruction:
You are a customer-help avatar for [COMPANY].
Your job:
- Help users with [ALLOWED TASKS].
- Ask for only the minimum information needed.
- Use the approved tools only when required.
- Explain what you are doing before a tool call.
- If a tool returns no result, say that clearly.
- Never invent an order status, price, policy or account fact.
- Escalate [PAYMENTS / COMPLAINTS / SAFETY / LEGAL REQUESTS] to a human.
- Tell the user that you are an AI assistant at the start of the session.
- Do not claim to be a real person.
A narrow role is easier to test than a general “help with anything” prompt. Define prohibited actions and escalation conditions before adding more capabilities.
4. Test turn-taking and interruptions
Run short tests for:
- normal question and answer;
- the user interrupting halfway through a response;
- silence and background chatter;
- a switch between two supported languages;
- camera input that is unclear or irrelevant;
- a failed or slow tool call;
- a request outside the avatar’s role.
Gemini 3.8 Live enables affective dialogue and proactive audio by default, and Google says its updated interruption handling waits rather than talking over a user in relevant tool-response situations.[3]
5. Connect one read-only tool
Start with a low-risk, read-only function such as order lookup or appointment availability. Do not make the first integration a payment, refund, deletion or account-change tool.
Google’s developer guide supports non-blocking asynchronous calls so the agent can acknowledge the request while the backend work continues. A blocking call can also be canceled when the user changes direction.[3]
6. Review every transcript and failure
For each test session, check:
- Did the avatar disclose that it was AI?
- Did it request unnecessary personal information?
- Did the spoken answer match the backend result?
- Did lip-sync or expression create a misleading impression?
- Did it keep talking after the user interrupted?
- Did language switching preserve names, numbers and policy terms?
- Did the tool execute exactly once?
- Was a human handoff available?
Only expand access after the narrow workflow passes repeatable tests.
Developer quickstart for the live model
Google’s official quickstart uses the Gen AI SDK, Application Default Credentials and the gemini-3.8-live model ID.[3] Install the current packages:
pip install --upgrade google-genai websockets numpy
Then establish a basic audio session:
import asyncio
from google import genai
from google.genai import types
PROJECT_ID = "YOUR_PROJECT_ID"
LOCATION = "us-central1"
MODEL_ID = "gemini-3.8-live"
client = genai.Client(
enterprise=True,
project=PROJECT_ID,
location=LOCATION,
)
config = types.LiveConnectConfig(
response_modalities=["AUDIO"],
system_instruction="You are a concise support assistant. Never invent account facts.",
)
async def main():
async with client.aio.live.connect(model=MODEL_ID, config=config) as session:
await session.send_realtime_input(text="Hello. What can you help me with?")
async for message in session.receive():
if message.server_content and message.server_content.output_transcription:
print(message.server_content.output_transcription.text, end="")
asyncio.run(main())
This establishes the live conversational foundation; it does not create a custom avatar by itself. Google’s Live Avatar documentation uses video response configuration and the enterprise avatar controls, while custom reference-based avatars remain allowlist-only.[2][3]
For production, follow the official Gemini Live API documentation for authentication, streaming, session management and tool-response handling.[4]
Practical deployment checklist
Conversation design
- State that the avatar is AI in the opening turn.
- Keep answers short enough for spoken interaction.
- Confirm names, dates, amounts and identifiers before acting.
- Give the user a clear way to stop, repeat or speak to a person.
- Never make silence look like consent.
Tool safety
- Use least-privilege credentials.
- Separate read actions from write actions.
- Require explicit confirmation before changes.
- Make tool calls idempotent where possible.
- Log the request, response, call ID and outcome without storing unnecessary personal data.
- Set timeouts and return a plain “no result” state instead of fabricating an answer.
Avatar and identity safety
- Use a prebuilt avatar unless a custom likeness is genuinely necessary.
- Obtain documented consent for a custom face and voice.
- Do not imitate public figures or imply endorsement.
- Keep an on-screen AI label visible.
- Review local privacy, biometric and recording requirements before deployment.
- Provide a way to report impersonation or harmful output.
Google says generated audio and video streams include SynthID watermarks, but watermarking does not replace user-facing disclosure, consent or operational controls.[1][2]
Quality assurance
- Test different microphones, cameras, browsers and network conditions.
- Check noisy rooms, accents and language switching.
- Test dropped connections and session recovery.
- Compare the transcript, spoken response and tool output.
- Validate that the avatar does not conceal delays with false statements.
- Red-team prompt injection through speech, camera text and shared screens.
Good use cases—and poor ones
Good first use cases are bounded and reversible: product navigation, FAQ help, appointment availability, guided onboarding, interactive training and read-only order status.
Poor first use cases include autonomous refunds, medical diagnosis, legal advice, unsupervised financial recommendations, identity verification based only on the avatar conversation, or any workflow where one wrong tool call causes irreversible harm.
For marketing, a live avatar may be useful when it helps visitors compare products or navigate a complex service. If the goal is a polished one-way promotional clip, an editor such as Google Vids with Gemini Omni is simpler and easier to review before publication.
FAQ
Is Gemini 3.8 Live Avatar available to consumers?
The launch is for Gemini Enterprise. Google directs users to Gemini Enterprise’s multimodal live studio and enterprise API documentation, not to a general consumer avatar creator.[1][2]
What is the Gemini 3.8 Live model ID?
The official developer guide lists gemini-3.8-live.[3]
Can I upload a photo to make a custom avatar?
Google says custom avatars can be created from a high-quality reference image, but custom avatar creation is currently restricted to enterprise allowlisting and verification. Use a prebuilt avatar unless your organization has approval.[1][2]
Does Live Avatar support multiple languages?
Google says Live Avatar can transition across 97 languages while adapting lip-sync and expressions. Test the exact languages, names and domain terms you need before launch.[1][2]
Can the avatar call APIs while it talks?
Yes. Gemini 3.8 Live supports non-blocking asynchronous tool calls, allowing the conversation to continue while a backend task runs.[1][3]
Is Live Avatar the same as Gemini text-to-speech?
No. TTS turns a prepared script into audio. Gemini 3.8 Live is a real-time conversational model that accepts continuous multimodal input, can use tools and can generate synchronized avatar video.[1][3]
Is the generated avatar video watermarked?
Google says the generated audio and video output carries imperceptible SynthID watermarking.[1][2]
Last checked: September 24, 2026.