NVIDIA Introduces PersonaPlex for Natural Full-Duplex AI Conversations

NVIDIA has unveiled PersonaPlex, a full-duplex conversational AI system that allows users to define both the voice and role of the AI while maintaining natural, real-time interactions, according to the company’s research publication.

PersonaPlex builds on the Moshi architecture from Kyutai and integrates a 7-billion-parameter model capable of listening and speaking simultaneously. This design removes the latency typical of cascaded systems that separate speech recognition, language generation, and speech synthesis. The model uses both text and voice prompts to define its conversational behavior, producing responses with natural timing, interruptions, and backchannel cues.

Training data for PersonaPlex combines real human conversations from the Fisher English corpus with synthetic dialogues generated using large language models and text-to-speech systems. This blend enables the model to exhibit realistic speech patterns while following role-specific instructions for scenarios such as customer service or teaching.

Evaluations on conversational benchmarks show that PersonaPlex outperforms other open and commercial systems in turn-taking, response latency, and task adherence. The model weights and code are available under open licenses, with the base Moshi model licensed under CC-BY-4.0 from Kyutai.