ElBruno.PersonaPlex 0.6.0

ElBruno.PersonaPlex

HuggingFace Models:

Run NVIDIA PersonaPlex-7B-v1 full-duplex speech-to-speech locally from C# using ONNX Runtime — no Python needed at inference time. Models are downloaded automatically on first run.

Pre-exported ONNX models are hosted on HuggingFace: elbruno/personaplex-7b-v1-onnx

Features

• Local Speech-to-Speech Inference — Run PersonaPlex entirely on your machine using ONNX Runtime
• Automatic Model Download — Models download from HuggingFace on first run
• Full-Duplex Capable — Based on NVIDIA's Moshi architecture for simultaneous listen + speak
• Persona Control — Text-based role prompts for customizable AI personas
• Voice Selection — Multiple pre-packaged voice embeddings (Natural + Variety voices)
• GPU Acceleration — Optional CUDA or DirectML support via SessionOptions injection
• Multi-Language — Supports conversational interactions in English
• Shared Model Cache — Models stored once in %LOCALAPPDATA%/ElBruno/PersonaPlex, shared across all apps

Quick Start

Install via NuGet

dotnet add package ElBruno.PersonaPlex

Basic usage in C#

using ElBruno.PersonaPlex.Pipeline;

// Models download automatically on first run
using var pipeline = await PersonaPlexPipeline.CreateAsync("models");

// Speech-to-speech with persona
await pipeline.ProcessAsync(
    inputAudioPath: "input.wav",
    voicePreset: VoicePreset.NATF2,
    textPrompt: "You are a friendly assistant.",
    outputPath: "output.wav");

GPU Acceleration

using ElBruno.PersonaPlex.Pipeline;

// CUDA (NVIDIA)
var pipeline = await PersonaPlexPipeline.CreateAsync(
    sessionOptionsFactory: SessionOptionsHelper.CreateCudaOptions);

// DirectML (any GPU on Windows)
var pipeline = await PersonaPlexPipeline.CreateAsync(
    sessionOptionsFactory: SessionOptionsHelper.CreateDirectMlOptions);

Supported Voices

PersonaPlex includes pre-packaged voice embeddings:

Real-Time Conversation Pipeline

For real-time audio conversations (VAD → STT → LLM → TTS), see the companion library:

👉 ElBruno.Realtime — Pluggable real-time audio conversation framework for .NET, following Microsoft.Extensions.AI patterns. Includes Whisper STT, QwenTTS, Silero VAD, and Blazor + Aspire samples.

Samples

See the Samples Guide for detailed instructions on each scenario, including what test content to use and expected output.

Run a Sample

cd src/samples/scenario-01-simple
dotnet run -- myrecording.wav

Documentation

Python Tools

The python/ directory contains tools for exporting ONNX models from PyTorch weights. These are only needed if you want to re-export or customize models — they are not required for running the C# pipeline.

Exported ONNX Models

The Mimi encoder/decoder models are hosted on HuggingFace and auto-downloaded by the C# library.

The LM backbone (7B parameters) can be exported locally using:

cd python
pip install -r requirements.txt
python export_onnx.py --component lm --output onnx_exports/ --moshi-path /path/to/personaplex/moshi --quantize fp16

Note: The onnxruntime-genai model builder does NOT support the PersonaPlex/Moshi architecture (it's not a standard HuggingFace model type). The manual export via export_onnx.py works and produces a validated, loadable ONNX model. FP16 inference on CPU takes ~0.4s per step with ~25s model load time.

Building from Source

git clone https://github.com/elbruno/ElBruno.PersonaPlex.git
cd ElBruno.PersonaPlex
dotnet build
dotnet test

Requirements

• .NET 8.0 or .NET 10.0 SDK
• ONNX Runtime compatible platform (Windows, Linux, macOS)
• Sufficient disk space for model files

Contributing

Contributions are welcome! Here's how to get started:

1. Fork the repository
2. Create a branch for your feature or fix: git checkout -b feature/my-feature
3. Make your changes and ensure the solution builds: dotnet build
4. Run tests: dotnet test
5. Submit a pull request with a clear description of the changes

Please open an issue first for major changes or new features to discuss the approach.

References

• NVIDIA PersonaPlex GitHub
• Original model (PyTorch)
• PersonaPlex Paper
• Pre-exported ONNX models

👋 About the Author

Hi! I'm ElBruno 🧡, a passionate developer and content creator exploring AI, .NET, and modern development practices.

Made with ❤️ by ElBruno

If you like this project, consider following my work across platforms:

• 📻 Podcast: No Tienen Nombre — Spanish-language episodes on AI, development, and tech culture
• 💻 Blog: ElBruno.com — Deep dives on embeddings, RAG, .NET, and local AI
• 📺 YouTube: youtube.com/elbruno — Demos, tutorials, and live coding
• 🔗 LinkedIn: @elbruno — Professional updates and insights
• 𝕏 Twitter: @elbruno — Quick tips, releases, and tech news

License

This project is licensed under the MIT License — see the LICENSE file for details.

Related Projects

• ElBruno.Realtime — Real-time audio conversation framework (STT, TTS, VAD)
• ElBruno.QwenTTS
• ElBruno.VibeVoiceTTS
• ElBruno.Text2Image
• ElBruno.HuggingFace.Downloader