Word recognition and automated epenthesis removal for Indonesian sign system sentence gestures

This research <span>focuses on building a system to translate continuous Indonesian sign (SIBI) gestures into text. In gesture, signer will add an epenthesis (transitional) which is hand movement with no meaning but needed connect the of one word next in gesture. Reducing number irrelevant inputs model through automated removal can improve system's ability recognize words gestures. We implemented threshold conditional random fields (TCRF) identify The dataset consists 2,255 videos representing 28 common sentences SIBI. translation MobileNetV2 as feature extraction technique, removing found by TCRF, and long short-term memory (LSTM) for classifier. With MobileNetV2-TCRF-bidirectional LSTM model, best error rate (WER) sentence accuracy (SAcc) were 33.4% 16.2%, respectively. Intermediate-stage processing steps consisting sandwiched majority voting TCRF labels whose frames less than two frames, along output grouping, able reduce WER from 3.4% increase SAcc 16.2% 80.2%.</span>