Comparing the Coordinate-Sequencing for Human's Joints Models on Thai Finger Spelling

Video-based sign language recognition plays a crucial role in enhancing communication and accessibility for those who are deaf or hard of hearing. However, developing and maintaining high-quality sign language datasets from video input is difficult, especially in Thai, because of the lack of a standard Thai finger spelling video dataset. To overcome this challenge, this article must focus on accumulating a larger total of 42 primary alphabets in Thai Finger Spelling from 22 signers with various appearances and genders. We conduct four deep learning-based archi-tectures: a sequence of coordinates of joints in human's body using LSTM, BiLSTM, GRU, and Transformer models. Indeed, we compare their performance under three scenarios of single-hand posture: single-stroke, two-stroke and three-stroke poses. Our results reveal that Transformer has the greatest performance on both in-sample and out-of-sample test sets.