Temporal window analysis for 1D-CNN in skeleton-based continuous sign language recognition

This paper proposes a skeleton-based continuous sign language recognition method applying 1D-CNN to MediaPipe features to address background sensitivity in 2D-CNN and limited temporal modeling in raw skeleton methods. We investigate the impact of temporal window widths (fixed vs. multi-scale), and compare LSTM and Transformer encoders. Evaluation on a standard dataset, RWTH-PHOENIX-Weather 2014T, demonstrates that a fixed 20-frame window with LSTM achieves the best performance with a Word Error Rate (WER) of 0.4134. These results indicate that appropriate temporal window selection is crucial and that LSTM outperforms Transformer for this task.