Beatrice Trainer for Windows

Train your own voice for the Beatrice voice changer, on your Windows PC.
Add recordings of one speaker, press Start, then import the finished model into Beatrice with Upload Custom Model.

Website · Beatrice for Windows · Beatrice for Mac · Trainer for Mac · Report a bug

---

✨ Why Beatrice Trainer

• 🎙️ A real Beatrice model of your voice. It runs the upstream Beatrice trainer on your recordings. There is no fake or "demo" training mode: if the trainer can't run, you see why.
• ⚡ Uses your NVIDIA GPU. The start script installs the CUDA build of PyTorch when it finds an NVIDIA card, and auto-tuning picks the batch size from your VRAM.
• 📥 Three ways to add audio. Drag in WAV or ZIP files, or import a dataset from Hugging Face or Kaggle.
• 📈 Watch it learn. Live step count, loss chart, time left and the trainer's own console output, in a browser tab that keeps working while you switch pages.
• 📦 One click to Beatrice. Download the finished model as a ZIP and import it into the voice changer with Upload Custom Model.
• ☁️ No fast GPU? Ready-made Colab and Kaggle notebooks train in the cloud for free.

📸 Tour

Training Progress, loss and the trainer's console output, updated every two seconds.	Dataset Name the speaker and add audio from your disk, Hugging Face or Kaggle.
Models Download a voice model as a ZIP for the voice changer, or keep the checkpoints.	Cloud training Open the notebooks in Colab or Kaggle and package your dataset for upload.
Light mode Dark, light or follow the system.	Honest errors If training can't start or crashes, the reason and a fix are shown right away.

_{Screenshots and the demo use sample data served by a local mock of the trainer API.}

🚀 Get started

1. Install Python 3.12 from python.org (3.10 to 3.13 work). In the installer, tick Add python.exe to PATH.
2. Download the trainer: download the ZIP and extract it, or clone it:

   git clone https://github.com/Project-Beatrice-V2/Beatrice-trainer-windows.git
   cd Beatrice-trainer-windows

3. Double-click start_windows.bat (or run it in Command Prompt). start.bat does the same.
4. Your browser opens http://127.0.0.1:8000. Next time, just run start_windows.bat again.

start_windows.bat creates a venv, installs PyTorch (the CUDA build when nvidia-smi finds an NVIDIA GPU, the cu128 wheels for RTX 50-series cards) and requirements.txt, downloads the trainer's pretrained weights from Hugging Face with python server.py --setup, then starts the server. The first run downloads about 3 GB, mostly PyTorch. If anything fails, the window stays open and shows the error.

Optional environment variables: BEATRICE_TORCH_INDEX=cu126|cu128|cpu forces a PyTorch wheel index, BEATRICE_NO_BROWSER=1 stops it opening a browser tab.

Note

An NVIDIA GPU with CUDA is recommended. The upstream trainer notes that its default settings use about 9 GB of VRAM and finish in about 40 minutes on an RTX 4090; auto-tune lowers the batch size on smaller cards. Without an NVIDIA GPU, training runs on the CPU: it works, but it is very slow. The cloud notebooks are a free alternative.

🧩 How training works

flowchart LR
  rec([🎙️ Recordings]) --> ds
  subgraph trainer [Beatrice Trainer · 127.0.0.1:8000]
    ds[Dataset<br/>beatrice-trainer/data/&lt;speaker&gt;] --> train[Training<br/>beatrice_trainer on CUDA or CPU]
    train --> out[Model folder<br/>paraphernalia_*]
  end
  out -- Download ZIP --> zip([model .zip])
  zip -- Upload Custom Model --> app([Beatrice voice changer])

Loading

1. Dataset. Enter a speaker name and add WAV files (ZIPs are unpacked), or import a Hugging Face or Kaggle dataset. Files go to beatrice-trainer/data/<speaker>/.
2. Train. Pick the speaker on the Training page. Auto-tune fills in batch size, gradient accumulation, workers and mixed precision for your PC; change them under Advanced settings. Press Start training.
3. Export. When the run finishes, open Models and click Download ZIP next to the voice model. Checkpoints (.pt.gz) stay in beatrice-trainer/output/<speaker>/.
4. Use it. In the Beatrice voice changer, open the model menu, choose Upload Custom Model and pick the ZIP.

💻 Requirements

Windows	Windows 10 or 11, 64-bit
GPU	NVIDIA GPU with CUDA recommended. CPU-only training works but is very slow.
VRAM	Upstream defaults use about 9 GB; auto-tune picks a smaller batch on 8 to 12 GB cards
Python	3.10 to 3.13 (3.12 recommended)
Disk	About 3 GB for PyTorch and packages, plus the trainer weights (about 175 MB), datasets and checkpoints

Troubleshooting

start_windows.bat closes immediately (fixed). Older versions of the script used characters that cmd.exe can't parse inside echo lines and if blocks, so the window could vanish before showing anything. The current script avoids them and pauses on every error. Update (git pull or download the ZIP again). If a window still closes, open Command Prompt in the folder and run start_windows.bat there to read the output.

Training finished instantly, or said "complete" straight away (fixed). Older versions quietly ran a stand-in "benchmark" loop when the real trainer couldn't start, so a run could look finished without training anything. That fallback is gone: the real trainer runs, and if it can't, the Training page shows Training could not start or Training failed with the reason. Update and run start_windows.bat again.

"The Beatrice trainer is not ready". Weights or packages are missing. Run start_windows.bat again, or open Settings and click Download or update trainer.

Training runs on the CPU although you have an NVIDIA GPU. Update the NVIDIA driver so nvidia-smi works, delete the venv folder and run start_windows.bat again. To force a wheel index, run set BEATRICE_TORCH_INDEX=cu126 (or cu128 for RTX 50-series) first.

"Python … is too old". Install Python 3.12 from python.org with Add python.exe to PATH ticked, delete the venv folder and run start_windows.bat again.

Out of memory during training. Lower Batch size under Advanced settings on the Training page, then start again.

The browser didn't open, or port 8000 is busy. Open http://127.0.0.1:8000 yourself. To use another port, run set BEATRICE_PORT=8001 before start_windows.bat.

Project layout

server.py           FastAPI server: API and the web UI on 127.0.0.1:8000
index.html          Web UI layout
index.css           The Aurora Void design system (dark and light)
app.js              Web UI logic: datasets, training, models, settings
fonts/              Geist and Geist Mono (SIL OFL)
start_windows.bat   Sets up the venv, installs PyTorch and packages, downloads weights, starts the server
start.bat           Same as start_windows.bat
start.sh            The same steps for macOS and Linux
beatrice-trainer/   The Beatrice trainer (fierce-cats/beatrice-trainer) plus webui_runner.py
colab_repo/         The Colab and Kaggle notebooks

🌐 The Beatrice family

🪟 Beatrice for Windows	The real-time voice changer that uses your models
🍎 Beatrice for Mac	The same app for Apple Silicon Macs
🌍 Website	Downloads, guides and voice samples
🧠 Trainer for Mac · Colab	Train on a Mac or in the cloud
📂 Dataset Web UI	Prepare recordings for training
🎭 Voice models	Ready-made models to import

🙏 Credits

• Beatrice trainer by Project Beatrice: fierce-cats/beatrice-trainer (MIT)
• Beatrice engine by prj-beatrice
• Geist and Geist Mono by Vercel (SIL Open Font License, see fonts/OFL-Geist.txt)

📜 License

The web UI and server are MIT licensed, © 2026 Project Beatrice V2 (see LICENSE). The Beatrice trainer in beatrice-trainer/ is MIT licensed by Project Beatrice; the fonts are under the SIL Open Font License.

If you trained a voice you love, ⭐ star the repo, it really helps.

Back to top ↑