PersonaPlex is a full-duplex conversational AI model that enables natural conversations with customizable voices and roles. It breaks the trade-off between traditional systems that allow customization but feel robotic and full-duplex models that feel natural but lock users into fixed voices and roles.

PersonaPlex offers full-duplex capability, allowing it to listen and speak simultaneously. It provides customizable voices and roles through text prompts, handles interruptions and backchannels naturally, and maintains consistent personas throughout conversations. The model demonstrates strong instruction following, empathy, accent control, and registration of user details from speech.

PersonaPlex uses a hybrid prompting architecture with two inputs: voice prompts capturing vocal characteristics and text prompts describing roles and context. It's built on the Moshi architecture with 7 billion parameters, using Mimi speech encoder/decoder and temporal/depth transformers. The dual-stream configuration enables concurrent listening and speaking for natural conversational dynamics.

The model delivers natural conversational experiences with low latency and authentic rhythm. Benefits include realistic customer service interactions, personalized assistant conversations, and adaptable role-playing scenarios. It handles diverse use cases from banking support to medical office reception and emergency scenarios.

PersonaPlex targets developers and researchers working on conversational AI applications. It integrates with existing AI workflows through open-source code and model weights. The technical foundation includes Helium language model for semantic understanding and generalization to out-of-distribution scenarios.