aiD Project — AI Sign-Language Avatars and the Elementary23 Curriculum Dataset for Deaf Students in Cyprus

The 'Artificial Intelligence for the Deaf' (aiD) project was coordinated by Associate Professor Sotirios Chatzis at the Cyprus University of Technology in Limassol, with a 32-researcher consortium that included Georgia Tech, the Hellenic Federation of the Deaf and the European Union of the Deaf. It ran from December 2019 to November 2023 under the EU's Marie Skłodowska-Curie Actions (Horizon 2020, grant 872139), with funding of €1.587m and a total project value reported up to €1.7m. It was motivated by evidence that over 60% of deaf children in Greece and Spain have severely limited reading and writing skills, which limits their access to spoken/written classroom content.

The project aimed to use deep learning and augmented reality to give deaf and hard-of-hearing students access to spoken classroom content in sign language. It set out to build a generative-AI app that produces sign-language video from text or speech in near real time via a personal avatar, plus a companion tool that overlays a signing avatar onto existing video, and to construct a Greek Sign Language dataset to train and evaluate these models.

The consortium built Elementary23, a Greek Sign Language dataset based on the official Greek elementary-school syllabus across six subjects (Greek Language, Mathematics, Religion, Environmental Study, History and Anthology), comprising 29,653 video-translation pairs from nine professional signers. A stochastic LWTA-Transformer model was trained on a curated 8,372-pair subset of the dataset and benchmarked against other sign-language translation datasets. The companion transcription/overlay tool was reported to have a memory footprint up to 70% smaller than prior solutions.

On the curated 8,372-pair Elementary23 subset, the stochastic LWTA-Transformer model reached BLEU-4 scores of 6.67 on validation and 5.69 on test data, ahead of the SWISSTXT-NEWS benchmark (0.41-0.46) and comparable to OpenASL (6.57-6.72). These are technical translation-quality metrics for the model, not measured classroom learning outcomes for deaf students.

As of the most recent public reporting, the team was still developing a commercial subscription model and no evidence of routine classroom deployment was found. The evaluation on Evidoria notes the project scores well on equity and inclusion given its focus on the deaf-education literacy gap, but rates evidence of learning impact, teacher-capability building and scalability/sustainability lower, since results reported are translation-quality benchmarks rather than classroom outcomes and the technology remains at prototype stage.