DeepMind unveils Genie 4, a world model that generates audio

Google DeepMind unveiled Genie 4 on Tuesday, a world model: an AI system that builds an interactive environment frame by frame and adjusts it to what the user does. It is the first of the company's models that generates audio natively while running in real time, DeepMind says.

Most people cannot try it yet. Genie 4 is a research preview for trusted testers only, and Google has not given countries, a price or a date. Google AI Ultra subscribers will get it "soon" in Project Genie, and early-access partners will receive it in the Gemini API and Google AI Studio, the company said.

Each world generates ambient sound, music and interactive dialogue on the fly. Users can change a scene with written prompts: move the camera, change the weather or add characters. Gemini Omni anchors how the world looks and behaves, and other Gemini models keep scenes consistent from frame to frame.

All video and audio Genie 4 produces will carry SynthID, Google's imperceptible watermark for identifying AI-generated content.

Google is the only source for these capabilities, and none has been independently tested. The company did not publish a maximum session length, resolution or frame rate. DeepMind itself acknowledges that visual details and object permanence can drift over time, especially in long sessions or after several edits. Complex physics, legible text and real-time audio accuracy remain open research problems.

Google points to two uses: training robots and agents in safe simulations, and creative tools for film, journalism, video games and manufacturing. For anyone working in games or video, the shift that matters is that world models are starting to generate voice and sound along with the picture, which makes SynthID more important for telling synthetic content apart.

Anyone who wants to try it can join the Project Genie waitlist.