<![CDATA[Communication barriers between deaf and hearing individuals remain a major challenge in inclusive education and social interaction. Most Indonesian Sign Language (Sistem Isyarat Bahasa Indonesia = SIBI) recognition systems focus solely on isolated gesture classification, without contextual understanding. This study proposes a real-time SIBI recognition system integrating LSTM-based temporal modeling and IndoBERT contextual language modeling. Hand keypoints were extracted from gesture sequences and processed using Long Short-Term Memory (LSTM) to recognize dynamic alphabet gestures. The predicted letter sequences were refined using a Trie-based lexical filter and IndoBERT to generate contextually appropriate word predictions. The dataset consisted of 780 gesture sequences representing 26 alphabet classes. The results showed that the system achieved 89.42% accuracy, 91.57% precision, 89.42% recall, and 88.74% F1-score while maintaining real-time performance at an average processing speed of 249.18 FPS and a latency of 129.46 ms per sample. Statistical evaluation confirmed that camera distance significantly affected recognition performance, while integrating IndoBERT improved contextual word-prediction accuracy from 18.2% to 58.64% (3.2× improvement). McNemar analysis further verified that the improvement was statistically significant, highlighting the effectiveness of contextual language modeling for enhancing semantic interpretation SIBI recognition. The proposed framework improves semantic interpretation and real-time usability for SIBI-based assistive communication systems.]]>
Copyrights © 2026
