Sevue

Sevue is a desktop application that turns live sign input into subtitle-like text over video and publishes the result to a virtual camera.

The app is built with PySide6, OpenCV, MediaPipe, and pyvirtualcam. It runs locally on your machine and is designed for low-latency, always-on desktop use.

What the App Does

• Captures frames from a selected physical camera
• Runs MediaPipe gesture recognition in a worker thread
• Builds short text output from recognized gestures
• Renders text stream
• Sends the processed stream to a virtual camera device
• Provides a desktop UI for camera/model/settings/shortcuts

installation

quick install

1. Grab the proper file for your OS from the Releases page.
2. Install the file like normal for your OS. Usually just double-click on the file.
3. That's it, sevue will be installed to your system like normal and shows up on the doc / start menu

Install From Source

Requirements

• Python 3.10 to Python 3.12
• A working physical camera or apps like DroidCam
• OS support:
  • Windows
  • Linux
  • macOS

steps

clone the project:

git clone https://github.com/codeweevers/sevue.git
cd sevue

create and activate a virtual env. For example with conda:

conda create -n sevue python=3.12
conda activate sevue

install requirements:

pip install --upgrade pip
pip install -r requirements.txt

Run

python sevue.pyw

Basic Usage

1. Launch Sevue.
2. Click Start Sevue.
3. In your conferencing/recording app, choose Sevue-VirtualCam (or your configured virtual cam target).

Runtime Architecture

• sevue.pyw: app entrypoint and single-instance lock/activation server
• controllers/main_window_controller.py: main orchestration (UI state, workers, tray, camera/model selection)
• models/state_model.py: persisted app state, settings, shortcuts, model registry, selected camera/model
• models/frame_buffer.py: shared frame buffer between camera and AI workers
• workers/threads.py:
  • CameraThread: capture, overlay rendering, virtual camera output
  • AIThread: gesture inference and subtitle text generation
• workers/camera_utils.py: cross-platform camera discovery and metadata
• services/model_registry_service.py: model import/validation/registry management
• services/startup_service.py: start-on-login integration (Windows/Linux)
• views/: Home/Settings pages and UI widgets

Configuration and Persistence

Sevue saves:

• selected camera (UID + resolved index)
• selected model + model registry
• app toggles (preview, flips, tray behavior, auto-start camera, etc.)
• keyboard shortcuts

Config location:

• Installed/frozen build: platform user config directory (platformdirs)
• Source/dev run: data/config.json in project root

Models

• Default runtime model is resolved via services/model_registry_service.py
• Additional .task models can be imported from Settings
• Imported models are copied into the app model storage directory and registered by name

Keyboard Shortcuts

Default configurable shortcuts:

• Ctrl+Shift+S start/stop camera
• Ctrl+Shift+M hide/show window
• Ctrl+Shift+C flip camera
• Ctrl+Shift+O flip subtitles
• Ctrl+Shift+H flip hand labels
• Ctrl+Shift+D toggle hand debug

Also supported:

• Esc window hide/show behavior

Packaging

PyInstaller spec is included at:

• train_installer_gen/sevue.spec

Windows installer-related assets/scripts are in:

• train_installer_gen/Install_SevueCam.bat
• train_installer_gen/Uninstall_SevueCam.bat
• train_installer_gen/installer scrypt.iss

Training (Optional)

Training can only be done on Linux and must use Python 3.10 or 3.11. This process is separate from the runtime app and is intended for creating and exporting custom gesture models.

1. Set up a compatible environment (Linux, Python 3.10‑3.11) and install the required package:

   pip install mediapipe-model-maker

2. Use the training script located at train_installer_gen/train.py to start model training. Note: edit the script to set the folder path of your dataset folder. The dataset should be structured as folders of images, where each folder name represents the gesture label (e.g., thumbs_up/, wave/, etc.).
3. Before or after training, run make_none.py to generate the required "none" files; you will need to edit that script to point the dataset folder path to your own data.
4. When the process completes the exported model will be written to exported_model/gesture_recognizer.task.
5. Import the resulting .task file using the choose model option in the app's settings.

This is separate from the runtime app and intended for creating/exporting gesture models.

Troubleshooting

• Camera cannot start:
  • Close other apps that may own the camera.
  • Re-open Settings and select the correct device.
• No virtual camera in external apps:
  • Verify your virtual camera driver/setup is installed and active.
• App already running:
  • Sevue is single-instance; launching again activates the existing window.

License

See LICENSE.