unknown

NVIDIA released PersonaPlex (Persona), an open-source full-duplex conversational AI model with near-zero latency. Unlike traditional voice assistants that process speech-to-text-to-speech sequentially, Persona uses a single end-to-end model that listens and speaks simultaneously, enabling natural back-channeling and active listening. Built on the Moshi architecture with 7 billion parameters, it was trained on 1,200 hours of real human conversations and 2,000 hours of synthetic role-specific data. The model requires significant GPU resources (24GB+ VRAM) and can be deployed via RunPod. While impressive in latency and conversational flow, it can become inaccurate and go off-rails when confused.

NVIDIA's New AI Voice Model Is INSANE! (PersonaPlex)

Development tricks and tips. Radically better observability stack. Ship higher-quality software faster. Be the hero of your engineering teams.