Skip to content
Get packs

PersonaPlex

Open-source speech-to-speech model that talks in real time and can adopt a set role and voice.

GitHub Download source
Stars
12k
Forks
1.6k
Licence
MIT
Language
Python
Last update
8 months ago
Type
AI model / agent

What it is

PersonaPlex is a real-time conversational speech model: you talk to it and it answers with speech, handling both sides of the conversation at once. You can steer it with a text prompt that sets its role (such as a teacher or a customer service agent) and pick a voice from a fixed set of pre-packaged options. It is built on the Moshi architecture and runs as a local web server or from the command line on audio files.

  • Run a live voice conversation with the model through the local web UI
  • Use a text prompt to make the model act as a customer service agent for a fictional business
  • Process an input WAV file offline and produce an output WAV and text transcript
  • Choose from 18 pre-packaged voices to change how the model sounds

Great for

  • Developers experimenting with voice assistants that need a set persona
  • Researchers evaluating full-duplex speech conversation
  • Teams prototyping customer service voice agents with custom role prompts
  • Anyone with a suitable NVIDIA GPU who wants to run a speech model locally

Not ideal for

  • People without a GPU with enough memory, since the README mentions memory limits and offers only a CPU offload option that needs extra packages
  • Users who need a polished end-user app, since the README describes a research code release and a basic web UI
  • Anyone who cannot accept the separate NVIDIA model license needed to download the weights
Step-by-step guide: how to use PersonaPlexInstall, first run, examples, tips and FAQs

How to get it

  1. 1

    Clone and build from source

    You'll need: Python · pip · PyTorch with GPU support · Opus audio codec development library (libopus-dev or opus-devel) · Hugging Face account and access token

    Install Opus codec (Ubuntu/Debian)

    sudo apt install libopus-dev

    Install the package from the repository

    pip install moshi/.

    Set Hugging Face token

    export HF_TOKEN=<YOUR_HUGGINGFACE_TOKEN>

    Launch the server

    SSL_DIR=$(mktemp -d); python -m moshi.server --ssl "$SSL_DIR"
  2. 2

    Run it

    • Install the Opus codec library and the moshi package.
    • Accept the PersonaPlex model license on Hugging Face and set HF_TOKEN.
    • Launch the server with the command in the README.
    • Open the Web UI at localhost:8998 in a browser and start talking.
  3. 3

    Get the most out of it

    • Use --cpu-offload if GPU memory is limited (requires the accelerate package)
    • Try the README's example role prompts for customer service or casual conversation
    • Use the offline script with the listed voice prompts to produce repeatable output with a fixed seed
    • Experiment with out-of-distribution prompts, as the README encourages

Downloads

Source code (.zip)main branch · straight from GitHub

Strengths

  • Supports real-time, full-duplex speech so it can talk and listen at the same time
  • Persona and voice can be controlled through prompts and voice selection
  • Code is released under the MIT license

Trade-offs

  • Model weights use a separate NVIDIA Open Model license and require a Hugging Face token and license acceptance
  • README notes possible extra setup for Blackwell GPUs and needs sufficient GPU memory for the model
  • Project description is minimal and the README is partly truncated, so detail is limited

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.