PersonaPlex - Speech-to-Speech Conversion
PersonaPlex is NVIDIA's full-duplex speech-to-speech model for live voice conversations. A text prompt sets the agent's role and the facts it uses, and one of 18 preset voices sets how it sounds.
Overview
PersonaPlex is a 7B-parameter conversational speech model from NVIDIA, built on Kyutai's Moshi architecture and weights. It is full-duplex: it listens and speaks at the same time, so it can wait through a pause or drop in a short "yeah" or "oh okay" while the other person talks.
Two prompts shape each session. The text prompt gives the agent its role, background and the facts it should use. The voice prompt sets how it sounds, and on Wiro you pick it from 18 presets.
NVIDIA trained the model on 7,303 real conversations from the Fisher English corpus plus synthetic assistant and customer service dialogues. On FullDuplexBench, NVIDIA reports a takeover rate of 0.950 on the user interruption task and 0.908 on smooth turn taking.
What you can build
- Customer service lines where the agent speaks as a named employee and uses the hours, menu or policies in the prompt. NVIDIA's examples include a restaurant, a waste collection service and a drone rental shop.
- Reception desks that take down a caller's details, like NVIDIA's medical office example that records new patient information.
- Voice assistants with a fixed persona, such as NVIDIA's "wise and friendly teacher" prompt.
- Characters for games or training drills, like NVIDIA's astronaut asking for help with a failing reactor.
Inputs
- Your voice. You talk through your microphone for the length of the session, in English.
- Prompt. Plain text saying who the agent is. The default is "You are a helpful assistant." For a service role, follow the example prompts on this page: company, business type, agent name, then an "Information:" line with the facts. For open conversation, start with "You enjoy having a good conversation." and add a topic or persona.
- Voice. Eight Natural voices (four female, four male) with a conversational delivery and ten more expressive Variety voices (five female, five male). The default is Variety Female 2.
- Seed. An optional number under advanced settings, 0 by default.
Outputs
Spoken replies stream back in the chosen voice as 24 kHz mono audio, with a text transcript of what the agent said and what you said.
Recommended settings
- Put every fact the agent needs into the prompt: names, opening hours, menu items, policies, what to verify. It cannot look anything up during a session.
- Pick a Natural voice for service calls and a Variety voice when the agent plays a character.
- NVIDIA uses the bare "You enjoy having a good conversation." prompt for its pause, backchannel and turn-taking tests, and the teacher prompt for its interruption test.
Limitations
- English only.
- Wiro offers the 18 preset voices. You cannot upload a sample of your own voice.
- NVIDIA trained the model on assistant, customer service and open-ended casual conversations. Other roles, like the astronaut, rely on the model generalizing beyond that training.
- A session ends on its own after half a minute in which neither side speaks.
Safety & compliance
The weights are released under the NVIDIA Open Model License, and NVIDIA marks the model as ready for commercial use. The Moshi base model it builds on is licensed CC-BY-4.0 by Kyutai. NVIDIA advises testing it with data from your own use case before deployment. Let callers know they are talking to an AI voice.
Example prompts
Great starting points for PersonaPlex-Realtime.
API quick start
Run PersonaPlex-Realtime with a single API call.
{
"prompt": "You are a helpful assistant.",
"voice": "VARF2",
"tempText": 0.7,
"topkAudio": 250
}curl -X POST "https://api.wiro.ai/v1/Run/nvidia/personaplex-realtime" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "You are a helpful assistant.",
"voice": "VARF2",
"tempText": 0.7,
"topkAudio": 250
}
JSONOther models you might like
Explore similar audio & speech models.
