OmniSign field note · Fact-checked September 23, 2026
Building in a Country With No Infrastructure
OmniSign started with a model idea and became a data, community, and systems problem. This is the corrected account of what we built, what the evidence supports, and what remains research.
The first problem was definition, not accuracy
We began by looking for data suitable for Lebanese Sign Language recognition. We could not find a machine-readable corpus that matched our scope, and the language itself varied across regions and teaching traditions. During collection, we documented variation associated with 21 teaching traditions. That figure describes variation observed during collection, not formal school partnerships or standardized coverage.
That changed the engineering sequence. Before model selection, we needed collection workflows, labels, vocabulary boundaries, and community guidance. The dataset was not an input to the project; building it was part of the project.
The verified data boundary
- 40,000+ captured samplesCommunity-collected sign samples before augmentation.
- 50,000+ training samplesThe post-augmentation training set, not an additional raw dataset.
- Arabic alphabet and 30+ expressionsThe bounded vocabulary used in the current recognition scope.
- Experimental clinical vocabularyA small set exists; broader medical and clinical coverage remains planned work.
A separate public repository contains an 87,000-image third-party American Sign Language alphabet dataset used for an early baseline. That dataset is not the community-collected Lebanese Sign Language dataset and is separate from OmniSign's reported figures.
Infrastructure became an engineering variable
Lebanon's electricity, connectivity, and institutional constraints meant that collection and development could not assume a clean lab environment. These conditions made local processing, recoverable workflows, and low-latency interaction important design directions.
Early public prototypes run webcam inference locally. That is evidence of a local prototype path, not proof of a universal on-device deployment, a production latency benchmark, federated learning, or a completed privacy architecture for the full platform.
What the system demonstrates
- Landmark extractionMediaPipe converts camera input into spatial and temporal features.
- Specialized recognition pathsModels handle isolated signs, vocabulary, and bounded temporal sequences.
- Multimodel integrationAn application layer routes recognition paths and presents text and speech output.
- Active R&DOmniSign does not claim unrestricted continuous translation or general real-world accuracy.
A result needs its boundary
OmniSign reached 98% in a narrow controlled internal evaluation. The historical repositories do not retain a signed-off report defining a reproducible evaluation split, so the number should remain stated at that level and no broader.
Real-world results vary with lighting, camera quality, signer variation, signing speed, and vocabulary. The 98% result is not unrestricted translation accuracy, a production service-level measure, or proof of performance for unfamiliar users and environments.
The team and the public record
The original student team was Layth Ayache, Tayseer Laz, Abou Baker Hussien Al Khatib, and Nour El Hariri, supervised by Dr. Oussama Mustapha. Rami Kronbi contributed later to computer-vision work and additional data collection.
The original team received the First Prize — Public Choice Award at Lebanon's National Final Year Project Demo Day 2025. Rafik Hariri University's announcement is the external source for the award and original team attribution.
Public artifacts and current scope
The main application and community-collected Lebanese Sign Language dataset are not published as open source. The public repositories preserve three narrower engineering artifacts:
- an early dataset-collection and augmentation prototype;
- the separate third-party ASL alphabet baseline; and
- an early multi-model recognition prototype.
The canonical OmniSign case study defines the current project scope, metrics, limitations, and contributor attribution.