NVIDIA's open-source full-duplex conversational speech AI system that enables real-time voice interactions with controllable voice characteristics and conversational persona. Built on the Moshi speech foundation model with a Helium LLM backbone, PersonaPlex processes incoming audio and generates contextually appropriate spoken responses simultaneously — supporting natural interruptions and overlapping speech. It ships with 16 pre-configured voices (Natural and Variety categories), audio conditioning for custom voice profiles, and text-based role prompts for persona definition. Designed for customer service automation, educational tutoring, and companion AI, the system supports GPU acceleration with CPU offload for limited VRAM environments. Deployment options include Docker with a web UI, offline evaluation tools, and experimental out-of-distribution prompt testing. The codebase is MIT-licensed; model weights are released under NVIDIA's Open Model License.