NVIDIA PersonaPlex

NVIDIA PersonaPlex: Full-Duplex Conversation With Any Role and Voice

PersonaPlex is an open-weight, real-time speech-to-speech model from NVIDIA's ADLR research lab that listens and speaks at the same time. A text prompt sets the role and context, and an audio voice prompt sets the voice, so developers get natural turn-taking without being locked into one fixed persona.

Key Features

• Full-duplex conversation: listens and speaks at the same time, and learns when to pause, interrupt or backchannel ("uh-huh", "oh").

• Single model, no cascade: replaces separate speech recognition, language model and text-to-speech steps with one model that streams its reply as the user speaks.

• Text role prompts: natural-language prompts define the role, background information and conversation context, from a teacher to a named customer service agent with prices, hours and rules.

• Voice prompts: an audio embedding sets vocal characteristics, speaking style and prosody, with 18 pre-packaged natural and varied voices.

• Persona that holds: keeps the chosen role and voice consistent across a long conversation.

• Built on Moshi: 7 billion parameters on Kyutai's Moshi architecture and weights, with the Mimi neural audio codec at 24 kHz and the Helium language model.

• Real and synthetic training data: 1,217 hours of real Fisher English conversations plus synthetic assistant and customer service dialogues.

• Open release: weights on Hugging Face under the NVIDIA Open Model License, code under MIT, and a research paper.

• Flexible deployment: runs on an NVIDIA GPU, with a CPU offload option for smaller GPUs, a web UI server for live tests and an offline script for evaluation.

Use Cases

• Customer service agents: voice agents that follow a business prompt, verify identity and capture caller details, as in NVIDIA's banking and medical reception demos.

• Question-answering assistants: a spoken assistant that handles interruptions and answers in a natural rhythm.

• Full-duplex research: a strong open baseline for turn-taking, interruption and backchannel studies, evaluated on FullDuplexBench.

• Character and casual conversation: open-ended role-play and companion-style chat driven by a short persona prompt.

Getting Started

• Website: https://research.nvidia.com/labs/adlr/personaplex/

• Model weights: https://huggingface.co/nvidia/personaplex-7b-v1

• Code: https://github.com/NVIDIA/personaplex

• Paper: https://arxiv.org/abs/2602.06053

PersonaPlex gives developers an open full-duplex speech model that keeps Moshi's natural conversation timing while letting them choose the voice and the role.