A Continuous Indian Sign Language Production System from Isolated Signs

Indian Sign Language (ISL) is a low-resource language, and the scarcity of well-annotated continuous datasets significantly restricts the development of robust applications for sign language recognition, translation, and production. The lack of such resources limits research advancements and the deployment of real-world ISL-based technologies. The primary objectives of this paper are: 1) To develop a sign language production (SLP) system that translates spoken language sentences into Indian Sign Language while incorporating proper co-articulation for natural and fluent signing, 2) To create a Synthetic Indian Continuous Sign Language (SynICS) dataset, which can serve as a valuable resource for future ISL translation research. To evaluate the effectiveness of our approach, we employ an off-the-shelf sign language translation (SLT) model for back-translation, assessing whether the synthetic signs can be reliably converted back into meaningful text. We train this model on the SynICS dataset and evaluate real-life videos from the iSign dataset. Our results are highly promising, achieving a BLEU-4 score of 17.56, with the translated sentences semantically matching the ground truth. These findings demonstrate the potential of synthetic continuous ISL data, generated from isolated signs, to improve Indian SLT task. The SLP system and the generated dataset will be publicly available on https://cs.rkmvu.ac.in/~isl/sl_gen.