Nvidia: Natural Conversational AI with Any Role and Voice

a month ago

NVIDIA PersonaPlex is a full-duplex conversational AI model that allows customization of voice and role while maintaining natural conversation dynamics.
It handles interruptions, backchannels, and authentic conversational rhythm, making interactions feel genuinely human.
PersonaPlex uses a hybrid prompting architecture with voice and text prompts to define conversational behavior.
The model is built on the Moshi architecture with 7 billion parameters and operates at a 24kHz sample rate.
Training data includes real conversations from the Fisher English corpus and synthetic dialogues for assistant and customer service roles.
Key findings include efficient specialization from pretrained foundations, disentangled speech naturalness, and emergent generalization beyond training domains.
PersonaPlex outperforms other systems on conversational dynamics, latency, and task adherence in benchmarks like FullDuplexBench and ServiceDuplexBench.
The model's code and weights are released under MIT License and NVIDIA Open Model License, respectively.

Hasty Briefsbeta