OmniSign — Lebanese Sign Language translation
Translates Lebanese Sign Language in real time, on phones, on the web, and on offline devices.
My partComputer vision engineer in a team of five
In short
A real-time translation system for Lebanese Sign Language spanning camera input, visual recognition, language output, and deployment across mobile, web, and offline embedded environments.
- What it is
- A real-time Lebanese Sign Language translation system.
- Why it matters
- General sign-language datasets do not transfer to local vocabulary, signing patterns, or the communication settings in which the system will be used.
- My part
- I co-founded the project and was the team’s computer vision engineer. Layth Ayache led the AI and data work; Nour El Hariri, Tayseer Laz, and Abou Baker Hussien Al Khatib were on the team; Dr. Oussama Mustapha supervised.
Step by step
- 1
Capture
Camera input is the primary channel; signer, background, lighting, and camera differ between deployments.
- 2
Recognize
A visual model classifies signs; performance on a curated dataset is not the same as performance for a specific signer in a specific room.
- 3
Translate
Recognized signs are converted into language output for a listener in a shared context.
- 4
Deploy
Deployment targets include mobile, web, and offline embedded configurations, each with its own constraints.
Results
- 1st prize
- Public Choice
- National FYP Demo Day, 2025
- 95–97%
- Development accuracy
- on our own 300,000-image dataset
- ~45 FPS
- Speed
- real time
- 3
- Pilots
- two Beirut coffee shops and one church
For engineersHow it is built, its limits, and its keywords.
How it works
A recognition model consumes camera input and produces language output; the system is deployed across mobile, web, and offline embedded targets. Uncertainty handling and the option to request a repeated sign are treated as design decisions.
Limits
- The 95–97% accuracy was measured in development, on our own dataset.
- The team’s project page counts 40,000 collected sign samples, while my count is a 300,000-image dataset. They measure different things, samples versus images, so the two are not added together.
Keywords
- Lebanese Sign Language
- gesture recognition
- edge deployment
- accessibility