TransSLR beats sign language benchmark with 80.39% accuracy, +10.46% gain

Automated sign language recognition has long struggled with under-represented languages. Central African Sign Language (CASL) is a case in point: the only public benchmark, CASL-W60, had a best-reported accuracy of just 69.93%, and fine-tuning large pre-trained models like WLASL fails because of limited data and a large lexical/visual domain gap. To close this gap, researchers at the Deep Learning Indaba 2026 (accepted for oral presentation) introduce TransSLR, a lightweight Temporal Transformer Encoder trained entirely from scratch on 64-frame normalized pose sequences.

TransSLR operates on geometric keypoint representations rather than raw RGB video, removing reliance on visual appearance and enabling signer-independent generalization. The encoder-only design with average pooling and a classification head drastically cuts computational overhead, making it feasible for deployment in low-resource environments. On CASL-W60, TransSLR achieves a new state-of-the-art 80.39% accuracy — a +10.46% improvement over the previous best. Extensive experiments against RGB-based and multimodal baselines confirm its effectiveness. This work highlights that carefully designed lightweight architectures can outperform heavy pre-trained models for niche languages, opening the door to scalable sign language recognition in regions where data and compute are scarce.