WMSM: Hebrew Sign Language recognition research with IEEE publication and product follow-through.

Existing datasets for Israeli Sign Language (ISL) were extremely limited — only 2,342 source videos covered the target vocabulary. That scarcity made standard supervised training impractical without significant augmentation.

The three-stage augmentation pipeline addressed this by applying spatial transformations, temporal perturbations, and landmark noise injection to expand the training corpus to 252,712 high-quality samples. Each stage was designed to preserve the biomechanical plausibility of hand movements while introducing enough variation to prevent overfitting.

The key architectural insight is that spatial hand landmarks extracted via MediaPipe Hands carry sufficient discriminative information for accurate gesture recognition — full video frame processing is unnecessary. This dramatically reduces computational requirements, enabling the entire model to train in just 24 GPU-hours (a 99.93% reduction compared to Seq2Seq baselines). A custom loss function then maintains temporal sentence coherence during live video inference, preventing the fragmentation that standard classification losses introduce when translating continuous signing into discrete words and sentences.

The research was published at IEEE FLLM 2025 in Vienna on November 26, 2025, and was subsequently presented at academic conferences in Israel. Following publication, the WMSM model was integrated into Handibur, an iOS video chat application delivering real-time Hebrew Sign Language-to-text translation at 40 FPS — a project that secured 2nd place in a collegiate Shark Tank competition.