An open-source, end-to-end neural voice conversion & model training ecosystem built for ultra-low latency, quality, and cross-platform flexibility.
┌───────────────────────────┐ ┌───────────────────────────┐ ┌───────────────────────────┐
│ 1. Dataset Web UI │ ───► │ 2. Model Trainer │ ───► │ 3. Real-Time Changer │
│ • Audio Slicing & Clean │ │ • macOS (MPS) & Win (CUDA│ │ • Ultra-low latency VST │
│ • Quality Validation │ │ • Free Google Colab │ │ • Live Discord/OBS stream│
└───────────────────────────┘ └───────────────────────────┘ └───────────────────────────┘
| Module / Repository | Platform Support | Features & Description |
|---|---|---|
| 🌿 Beatrice-website |  |
Official modern web portal, interactive audio showcase, documentation hub, and multi-lingual portal for Project Beatrice V2. |
| 🎭 Beatrice-voice-models |  |
Official repository of pre-trained Beatrice neural voice weights and models (including Donald Trump 5,000-step checkpoint). |
| 🎙️ Beatrice-voicechanger-macos |  |
Real-time Beatrice voice changer for macOS featuring Metal/MPS hardware acceleration and ultra-low latency. |
| 🎙️ Beatrice-voicechanger-windows |  |
Fast, lightweight, low-latency live voice conversion running locally with GPU acceleration on Windows. |
| 🧠 Beatrice-trainer-macos | |
Train Beatrice voice models locally on macOS with a clean desktop UI and optimized Apple Silicon pipeline. |
| 🧠 Beatrice-trainer-windows |  |
Streamlined Windows trainer for custom voice models with complete offline support and efficient training. |
| ☁️ Beatrice-colab |  |
Cloud training notebooks for Google Colab & Kaggle. Train models on free T4/A100 GPUs without local hardware. |
| 📂 Beatrice-dataset-webui-macos |  |
Web-based dataset creator for macOS. Automatically slice long audio files, clean noise, and format training sets. |
| 📂 Beatrice-dataset-webui-windows |  |
Intuitive web UI for Windows to build, process, and package voice datasets for Beatrice model training. |
Step 1: Dataset Preparation (Web UI)
- Launch Beatrice-dataset-webui-macos or Beatrice-dataset-webui-windows depending on your OS.
- Load your raw voice recordings into the interface.
- Automatically slice long files into clean audio segments, validate transcriptions/silence, and export a ready-to-train dataset zip.
Step 2: Model Training
- macOS (Local): Use Beatrice-trainer-macos for native Apple Silicon MPS training.
- Windows (Local): Use Beatrice-trainer-windows for CUDA / DirectML hardware acceleration.
- Cloud (Free GPU): Launch Beatrice-colab on Google Colab or Kaggle to train without needing a high-end local GPU.
Step 3: Real-Time Live Inference
- Download pre-trained weights from Beatrice-voice-models or export your custom model.
- Open Beatrice-voicechanger-macos or Beatrice-voicechanger-windows.
- Route your physical microphone into the voice changer and output to a virtual audio channel (e.g. VB-Cable / BlackHole).
- Use transformed voices live in Discord, OBS, games, or live streaming apps!
Built with ❤️ by the Project Beatrice V2 Team • Released under the MIT License
