PersonaPlex is a full-duplex conversational AI model that breaks the trade-off between customization and naturalness in conversational AI. Traditional systems allow voice and role customization but feel robotic with awkward pauses and unnatural turn-taking, while full-duplex models provide natural conversations but lock users into fixed voices and roles. PersonaPlex delivers truly natural conversations while maintaining customizable personas throughout interactions.

PersonaPlex offers full-duplex capability, enabling it to listen and speak simultaneously for low-latency interaction. The model supports customizable voices through audio embeddings that capture vocal characteristics, speaking style, and prosody. It allows role definition through natural language text prompts describing the role, background information, and conversation context. PersonaPlex handles interruptions, backchannels, and authentic conversational rhythm while maintaining persona coherence throughout extended interactions.

The architecture uses a hybrid prompting system with voice prompts capturing vocal characteristics and text prompts defining conversational roles. Built on the Moshi architecture with 7 billion parameters, PersonaPlex uses Mimi speech encoder and decoder components operating at 24kHz sample rate. The dual-stream configuration allows concurrent listening and speaking, while the underlying Helium language model provides semantic understanding and generalization capabilities.

Benefits include natural conversational dynamics with realistic turn-taking and interruption handling, customizable personas for various roles including assistants, customer service agents, and fictional characters, and strong task adherence across different scenarios. The model demonstrates generalization beyond training domains, handling technical vocabulary and emotional urgency appropriate to different contexts.

Target users include developers building conversational AI applications, researchers working on speech AI systems, and organizations needing customer service or assistant solutions. The model integrates with existing AI workflows through open-source code and model weights available under MIT License and NVIDIA Open Model License respectively.