cultrix/personaplex repository overview

⁠PersonaPlex Docker Image

This Docker image provides a ready-to-run environment for PersonaPlex, NVIDIA's real-time, full-duplex speech-to-speech conversational AI model. Built on the Moshi architecture (7B parameters), it enables natural voice interactions with customizable personas (roles via text prompts) and voices (via audio embeddings). Ideal for deploying on GPU-accelerated platforms like RunPod, local Docker, or Kubernetes.

⁠Key Features

• Full-Duplex Conversation: Simultaneous listening and speaking with low latency.
• Persona & Voice Control: Text-based roles (e.g., teacher, astronaut) + pre-packaged voice embeddings (NATF0–NATF3, VARM0–VARM4).
• Web UI: Browser-based interface for live microphone/speaker interactions.
• Offline Evaluation: Process WAV inputs to generate responses and transcripts.
• CUDA Support: Optimized for CUDA 12.8, compatible with RTX 40-series and Blackwell GPUs (e.g., RTX 5090).
• Dependencies: Python 3.12, PyTorch (CUDA 12.8), and Hugging Face libraries.

⁠Model Details

• Repo: nvidia/personaplex-7b-v1⁠ (gated; requires HF token and license acceptance).
• Paper: PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models⁠.
• Base Framework: kyutai-labs/moshi⁠.

⁠Usage

⁠Quick Start

Pull the image:

docker pull yourdockerusername/personaplex:latest

Run with GPU (requires NVIDIA Docker runtime):

docker run --gpus all -p 8998:8998 -e HF_TOKEN=your_hf_token -e NO_TORCH_COMPILE=1 -v ./cache:/root/.cache yourdockerusername/personaplex:latest

• Access the Web UI at http://localhost:8998.
• First run downloads the model (~14GB; cached in volume).

⁠Environment Variables

• HF_TOKEN: Hugging Face access token (required for model download).
• NO_TORCH_COMPILE=1: Disables Torch compile for compatibility.

⁠Build from Source

Clone the repo: Teachmetech/personaplex-docker⁠.

docker build -t yourdockerusername/personaplex:latest .

⁠Requirements

• Docker with NVIDIA GPU support.
• GPU with ≥24GB VRAM recommended (e.g., RTX 4090+).
• Accept NVIDIA Open Model License on Hugging Face.

⁠Troubleshooting

• 401 Unauthorized: Ensure HF_TOKEN is set and license accepted.
• Model Download Time: 10–30+ minutes on first run.
• For RunPod deployment, see the RunPod Template README⁠ (adapt as needed).

⁠Credits & License

• Original: NVIDIA/personaplex⁠.
• Dockerized by: Teachmetech/personaplex-docker⁠.
• Model License: NVIDIA Open Model License.
• Code License: MIT.

For issues, check the GitHub repo or NVIDIA docs.