SignBridge is an open-source, in-browser bridge between sign language and text/speech, built so that deaf and hearing people can communicate without an interpreter or special hardware, just a webcam.

What it does

• Sign to text: tracks up to two hands with MediaPipe (21 landmarks per hand) and recognizes signs with a small k-nearest-neighbours classifier that runs entirely in the browser. Features are normalized (wrist as origin, scaled by palm length), so recognition does not depend on where the hand is or how far it is from the camera. A stability filter only commits a letter after the same sign is held for about 10 frames, so flicker does not type garbage.
• Train your own signs: record 30 samples per sign in about a second, and export or import the dataset as JSON so a team can build the alphabet together. No model training step, no server.
• Gemma cleanup: the raw transcript (for example HELLOWORLD) is sent to Gemma 4 through the Gemini API, which splits words and fixes casing and misread letters. If the API is slow or fails, a local formatter is used so the app never blocks.
• Speech to sign: spoken or typed text is converted into letter-by-letter sign cards.

Why

Most sign-language tools need cloud services or special cameras. SignBridge keeps hand tracking and recognition on the user’s device, and lets any community teach it its own signs.

Status

The sign-recognition engine and both UI panels are built. Recognition quality depends on the dataset the user records, and the shared A–Z dataset is still being collected. Future development will expand the platform beyond letter-level recognition to include speech-to-text conversion and word-level sign generation. This will allow users to enter a word and dynamically generate a corresponding sign language card displaying its visual representation (for example, entering “YES” generates the sign card for YES). This will extend SignBridge towards more natural, vocabulary-level communication rather than relying solely on individual alphabet letters.