Voice agents with Realtime Video — Sidney Primas, LemonSlice

Lemon Slice aims to pass the avatar Turing test by developing photorealistic, real-time AI avatars indistinguishable from humans. The company utilizes human-centric world models to capture complex behaviors, including micro-expressions, object interactions, and emotional responses. To achieve real-time performance, the system employs single-step denoising and a specialized architecture that prevents error accumulation during long-duration generation. Beyond visual fidelity, the core technical challenge lies in the "model harness," which orchestrates data streaming to ensure seamless, stutter-free interaction. Looking ahead, the focus shifts toward end-to-end "EQ" models capable of processing raw video and audio to manage internal emotional states, while separate "IQ" models handle logic and tool-calling tasks. This approach seeks to move beyond scripted interactions toward deeply responsive, emotionally aware digital entities that function as a visual layer for AI agents.