Sovereign Sign: An Edge-Native Framework for Privacy-Preserving Sign-to-Text Translation

This paper proposes Sovereign Sign, an end-to-end, on-device sign-to-text translation system designed to resolve the privacy-utility paradox in sign language processing, with particular attention to the socioeconomic and infrastructural realities of the Global South, specifically Ghana. The proposed pipeline converts live video into 3D landmark vectors via on-device pose and face tracking and immediately discards the raw frames, ensuring zero persistence of identifiable data. Calibrated Laplace or Gaussian noise is then added to each keypoint coordinate, achieving (ε, δ)-differential privacy and providing a formal, quantifiable privacy guarantee. A compact translation model, a Spatio-Temporal Graph Convolutional Network (ST-GCN) encoder followed by a Transformer decoder with lightweight LoRA adapter layers, runs efficiently on mid-range mobile hardware such as the Me-diaTek Dimensity 700 and Qualcomm Snapdragon 680, the SoCs that dominate the West African consumer market. Crucially, the system explicitly incorporates facial and mouth landmarks (non-manual features, NMFs) to preserve grammatical nuance in Ghanaian Sign Language (GhSL), a dimension entirely absent from prior work. Evaluation on both the ASL How2Sign corpus and the Ghanaian SignTalk-GSL dataset demonstrates that Sovereign Sign approaches cloud-level translation accuracy (BLEU-4 ≈ 28.15, WER 24.80%) while operating at ∼31 ms per frame under strict privacy constraints, consuming under 350 mAh/hr on commodity hardware.