TurkSign446: A Comprehensive Dataset for Turkish Sign Language Recognition

The TurkSign446 dataset is a large-scale Turkish Sign Language (TSL) dataset developed for research in sign language recognition. It includes 446 words (133 static, 313 dynamic), collected from 10 participants with 10 repetitions each, resulting in 44,600 videos (over 1.3 million frames). Each video was processed with MediaPipe Holistic, extracting 1,662 features per frame in NumPy format. The dataset was designed to be device, background, and resolution independent, providing rich representational capacity through hand, face, and body landmarks. This dataset was used in the study “TurkSign446: A Comprehensive Dataset and Hybrid Deep Learning Approach for Real-Time Turkish Sign Language Recognition”, where hybrid deep learning architectures (CNN, LSTM, GRU, Bi-LSTM) were evaluated. The best-performing model (CNN+GRU) achieved 97.88% accuracy in offline tests and 85.5% (user-dependent) / 80.27% (user-independent) in real-time tests with an inference time of 45–60 ms. Citation If you use this dataset or code, please cite: @article{Karaci2025_TurkSign446, title={TurkSign446: A Comprehensive Dataset and Hybrid Deep Learning Approach for Real-Time Turkish Sign Language Recognition}, author={Torun, Cumhur and Karacı, Abdulkadir}, journal={Pattern Analysis and Applications}, year={2026}, publisher={Springer} } The TurkSign446 dataset fills an essential gap in TSL research, offering a robust benchmark for developing real-time and user-independent sign language recognition systems.