TransSLR Sets CASL-W60 Sign Recognition Benchmark
Carnegie Mellon University Africa researchers released the TransSLR preprint on August 3, reporting 80.39% signer-independent accuracy on the CASL-W60 sign-language benchmark. The authors attribute the result to a lightweight temporal Transformer trained from scratch on normalized pose sequences, exceeding the previously reported 69.93% accuracy by 10.46 percentage points.
TransSLR, a lightweight Transformer for sign-language recognition, reached 80.39% signer-independent accuracy on the CASL-W60 benchmark, according to an August 3 arXiv preprint from researchers at Carnegie Mellon University Africa. The authors report that the result exceeds the benchmark's previous best reported accuracy of 69.93% by 10.46 percentage points.
The paper describes CASL as Central African Sign Language and presents the task as a low-resource recognition problem. According to the authors, fine-tuning models trained on higher-resource sign-language data did not close the performance gap, which they attribute to limited CASL training data and lexical and visual differences from large datasets such as WLASL. The paper is an arXiv preprint.
Pose sequences instead of RGB video
TransSLR uses a Temporal Transformer Encoder trained from scratch on 64-frame normalized pose sequences. The reported architecture applies average pooling and a classification head, rather than processing raw RGB video frames. The authors state that using geometric keypoint representations reduces dependence on visual appearance and supports evaluation on signers not seen during training.
The accompanying Hugging Face repository specifies an input tensor shape of (batch, 64, 225). Its documentation describes a linear 225-to-512 feature layer, temporal batch normalization, four Transformer layers with eight attention heads, and a two-layer MLP classifier with 0.5 dropout. It also states that input landmarks require mid-hip centering and shoulder-distance scaling.
The repository reports evaluation on 862 files from independent signers, with 80.39% overall accuracy and a 78.33% weighted F1 score across 60 sign classes. It provides downloadable PyTorch weights and model code, although no hosted inference provider is listed.
Data provenance needs clarification
There is a documentation inconsistency across the released materials. The arXiv paper identifies CASL as Central African Sign Language, while the Hugging Face repository labels its project as "Chinese-American Sign Language." The paper and repository share the TransSLR name and reported 80.39% accuracy, but practitioners evaluating the release should verify the dataset's language definition, licensing, participant protocol, and train-test split before treating the repository as a reproducible implementation of the preprint.
More broadly, low-resource sign-language systems can benefit from pose-based representations because they can reduce computational overhead and may avoid learning person-specific appearance cues. That does not by itself establish linguistic coverage or real-world accessibility: benchmark accuracy depends on the sign vocabulary, data collection conditions, annotation quality, and whether evaluation separates people across training and test sets. TransSLR's reported signer-independent result is therefore a useful benchmark result, subject to independent replication and clarification of the public release metadata.
Key Points
- 1TransSLR reports 80.39% signer-independent CASL-W60 accuracy, a 10.46-point improvement over the benchmark's prior reported best result.
- 2The model consumes normalized 64-frame pose sequences rather than RGB video, reducing visual processing requirements while targeting unseen-signer generalization.
- 3Conflicting CASL expansions in the paper and model repository make dataset provenance and release validation important for reproducibility-focused teams.
Scoring Rationale
The preprint reports a substantial benchmark gain for a low-resource sign-language recognition task and releases model artifacts that practitioners can inspect. Its significance is constrained by the narrow 60-class benchmark, preprint status, and inconsistent public metadata about CASL.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
