Griffin by Tavus

Reacting to pauses, expressions and gestures while talking in real time is what separates a natural call from a stilted one, so building a model specifically around face to face conversation makes sense. The 48 percent believed they were talking to a real person result is a striking number. How does it handle interruptions when both sides speak at once?

Griffin is Tavus’s first Human Interaction Model, built to understand and generate face-to-face interaction in real time.Problem: Most real-time AI systems work in turns. They wait for you to finish, process your speech, then respond. That can miss the pauses, expressions, interruptions, and visual context that make human conversations feel natural.Solution: Griffin is a full-duplex video-to-video model that sees, hears, and responds at the same time. It continuously evaluates the conversation and can speak, react, wait, or be interrupted without losing its place.What makes it different: Griffin combines perception, conversational decision-making, speech generation, and video generation into one real-time system. It can react to what it sees on camera while continuing to talk, generate gestures and expressions, and respond to the timing of a conversation.Key features:Sees and hears the user simultaneouslyFull-duplex conversationResponds at regular sub-second intervalsCan interrupt, backchannel, or be interruptedUnderstands visual contextModels expressions, gestures, tone, and behaviorGenerates the full scene, not just the faceGenerates 720p video in 320 ms chunksCan clone a speaker’s voice from about 10 seconds of audioGriffin-Lite is available as a research preview to select testersThe result: In Tavus’s live study, 48% of participants who had a one-minute video call with Griffin believed they had spoken with a real person. Tavus also reports Griffin ranked #1 on NVIDIA’s independent face-to-face AI test.Who it's for: Griffin is currently a research preview for selected early testers. Tavus says it is not available for customer use at this time and plans a wider release after addressing safety concerns.