Abstract
Skeleton-based sign language recognition (SLR) encodes body–hand keypoints rather than raw appearance, offering privacy and computational efficiency. This PRISMA-guided review synthesizes 19 studies (2015–2024) and contributes (i) a five-axis taxonomy of tasks, inputs/encodings, model families, evaluation protocols, and dataset anchors; (ii) a systematic table of study characteristics; and (iii) a quantitative meta-analysis that aggregates performance by model family. On curated isolated benchmarks, Top-1 accuracy frequently lies in the 93–98% range (e.g., CSL-500, AUTSL, LSM), whereas large-vocabulary WLASL subsets depress pose-only accuracy (≈63% on WLASL-100 and ≈24% on WLASL-2000). For continuous/online SLR, the single skeleton-based study with streamingdecoding reports word-error rates ≈11–19% (offline ≈9.5%). Across families, attention-augmented multi-stream GCNs and hybrid GCN+general-DNN models are consistently strong on curated isolated sets; TSSI+CNN is a portable alternative when holistic landmarks are available; and explicit 3D body–hand–face reconstruction yields stable gains on controlled subsets. Performance variability is driven primarily by vocabulary scale, articulator coverage, and whether skeletons are provided or inferred. We conclude that skeleton-only pipelines are mature for isolated recognition on curated corpora, while robust continuous recognition and large-vocabulary generalization will require broader multimodal datasets, standardized reporting, and tighter integration of non-manual articulators, with appearance+pose fusion likely beneficial in practice.