A bi-directional multilingual translation app for desktop and mobile, based on state-of-the-art real-time client side models.
- Desktop / Mobile ready UI
- Language selectors
- Translation direction selector
- Signed-to-spoken language translation
- Spoken-to-signed language translation
- Camera / File upload video inputs
- SignWriting hand shape and orientation estimation
- SignWriting facial features estimation
- Language identification (Detect Language) - TODO
- Segmentation - TODO
- Tokenization - TODO
- SignWriting to spoken language translation - TODO
- Text-to-speech
- Copy / share / edit translation - TODO
- Text input
- Microphone input - TODO
- Text-to-speech
- Spoken language text to SignWriting translation - TODO
- SignWriting to pose sequence - TODO
- Pose sequence to video - TODO (tensorflow/tfjs#5374)
- Text to pose sequence (server side)
- Rethink app icon
- MediaPipe doesn't work on mobile (google-ai-edge/mediapipe#1427)
This playground is intended for ease of prototyping real-time sign language models.
It includes, as a basic first step, MediaPipe Holistic pose estimation, on top of which other predictions are performed:
- Sign language detection (Model based)
- SignWriting hand orientation (Rule based)
- SignWriting hand shape (Model based)
- Partial SignWriting non-manuals - eyebrows, eyes, mouthing (Model based)