1
Load model
Press Load recognition model and wait for confirmation.
Live sign recognition
Hands and shoulders visibleCamera is off
Start the camera and allow permission.
Camera off
Position your hands inside the frame
Camera adjustments
Adjustments affect the preview and the image sent for landmark tracking. Actual torch support depends on the camera and browser.
◷
Ready
Start the camera or load the model.
ConnectionReady
Inference engineNot loaded
RecognitionWaiting
Features— / 86
Predicted class—
Confidence—
Inference time—
Processing FPS—
Camera starts only when requested. Review recognised text before relying on it in important conversations.
Your message
Editable textSpeech uses browser speech synthesis.
Runtime log
Initialising SignSpeak…
Quick start
Get started in four steps
Load your trained model, allow the camera, sign, then review the result.
2
Allow camera
Press Start camera and approve browser permission.
3
Sign clearly
Keep hands and shoulders visible. Hold a gesture steadily.
4
Review and speak
Check the message, edit if needed, and use Speak message.
Helpful tips
Get better recognition
Good lighting
Face a light source and avoid strong backlighting.
Hands visible
Keep hands, shoulders and upper body in frame.
Separate gestures
Release or move distinctly before repeating a sign.
Learn more
About SignSpeak
What signs are recognised?
Only classes included in the trained model's vocabulary. This is not a universal sign-language translator.
How does recognition work?
MediaPipe extracts hand, pose and face landmarks. The application builds an 86-feature representation and passes 20 frames to the GRU model using ONNX Runtime Web.
What are the limitations?
Results depend on lighting, camera position, tracking quality and training data. This is an experimental prototype.
NS
About the Developer
SignSpeak is developed by Naitik Srivastava, a B.Tech Electronics and Communication Engineering student interested in accessible technology, artificial intelligence, computer vision and embedded systems.