# AI use cases for PersonaPlex (NVIDIA/personaplex)

Repo: [NVIDIA/personaplex](https://github.com/NVIDIA/personaplex) | Use cases: 3 | Page: https://dedupedai.com/use-cases/nvidia-personaplex/

PersonaPlex is NVIDIA's real-time, full-duplex speech-to-speech conversational model based on Moshi that controls persona through text role prompts and audio voice conditioning.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Talk to a full-duplex voice model in your browser: https://dedupedai.com/use-cases/nvidia-personaplex/run-full-duplex-voice-agent-web-ui/
- Give a voice agent a role and a voice: https://dedupedai.com/use-cases/nvidia-personaplex/give-voice-agent-a-persona/
- Generate spoken replies from recorded input offline: https://dedupedai.com/use-cases/nvidia-personaplex/batch-generate-spoken-replies-offline/

## How to talk to a full-duplex voice model in your browser with PersonaPlex

Repo: [NVIDIA/personaplex](https://github.com/NVIDIA/personaplex) | Difficulty: Advanced | Time: 1 hour | Basis: From the README

Start the local server and hold a real-time spoken conversation through the web UI.

**Who it is for:** A developer evaluating natural, interruptible voice agents.

**The problem:** Turn-based voice bots feel unnatural.

**The outcome:** A local full-duplex conversation with low latency.

### Steps

1. Install Opus with `sudo apt install libopus-dev`
2. Install with `pip install moshi/.` and set `export HF_TOKEN=`
3. Start with `SSL_DIR=$(mktemp -d); python -m moshi.server --ssl "$SSL_DIR"`
4. Open the web UI address the server prints

### Prompt to paste into your agent

```text
Install NVIDIA/personaplex, start moshi.server with SSL, and hold a 1 minute conversation. Report latency observations and confirm the server ran on GPU.
```

### Caveat

Add --cpu-offload if the model does not fit in GPU memory.

Risk, credentials: The HF_TOKEN grants access to your Hugging Face account.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/nvidia-personaplex/run-full-duplex-voice-agent-web-ui/

## How to give a voice agent a role and a voice with PersonaPlex

Repo: [NVIDIA/personaplex](https://github.com/NVIDIA/personaplex) | Difficulty: Advanced | Time: 45 minutes | Basis: From the README

Combine a text role prompt with a voice prompt to keep a consistent persona.

**Who it is for:** A product team prototyping a customer service voice agent.

**The problem:** Voice agents drift in tone and role.

**The outcome:** Output audio and text that follow the given role and voice.

### Steps

1. Install PersonaPlex
2. Pick a voice prompt such as NATM1.pt
3. Write a role prompt in a text file
4. Run moshi.offline with --voice-prompt, --text-prompt and --input-wav

### Prompt to paste into your agent

```text
Run PersonaPlex offline with voice prompt <file>, role prompt <text> and input <wav>, then check output.json follows the role.
```

### Caveat

Blackwell GPUs need the extra PyTorch install step from the README.

Risk, voice cloning: Voice conditioning can imitate a voice; only use voices you have rights to.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/nvidia-personaplex/give-voice-agent-a-persona/

## How to generate spoken replies from recorded input offline with PersonaPlex

Repo: [NVIDIA/personaplex](https://github.com/NVIDIA/personaplex) | Difficulty: Advanced | Time: 30 minutes | Basis: From the README

Run offline inference on a WAV file to get the model's spoken reply and transcript.

**Who it is for:** A researcher building an evaluation set for voice agents.

**The problem:** Live sessions are hard to reproduce.

**The outcome:** Repeatable output.wav and output.json with a fixed seed.

### Steps

1. Install PersonaPlex and set HF_TOKEN
2. Prepare an input WAV
3. Run moshi.offline with --seed for reproducibility
4. Compare outputs across prompts

### Prompt to paste into your agent

```text
Run moshi.offline from NVIDIA/personaplex on <input.wav> with seed 42424242 and two different voice prompts. Report differences between the two outputs.
```

### Caveat

It is research code tied to specific model weights on Hugging Face.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/nvidia-personaplex/batch-generate-spoken-replies-offline/