AI‑Powered Assistive Translator for Deaf‑Mute Individuals

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Uluslararası Final Üniversitesi, Engineering, Software Engineering, Kıbrıs (Kktc)

Tezin Onay Tarihi: 2026

Tezin Dili: İngilizce

Öğrenci: Imane Boumedra

Danışman: Ghazaal Sheikhi

Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu

Özet:

Accessibility in cross-lingual deployment scenarios is limited because sign language translation (SLT) systems are typically trained for a single spoken-language target. This paper presents a novel two-stage framework that extends monolingual SLT optimisation to a cascaded multilingual text generation pipeline under realistic computational constraints. Building on a Sign Language Transformer baseline, we conduct systematic configuration-driven experimentation from baseline to optimised configuration for English and German over the PHOENIX-2014-T benchmark, identifying a high-performing setting (Optimised configuration–German) that achieves WER 31.98 and BLEU-4 21.23 on the German test set, and WER 48.39 and BLEU-4 20.88 on the English pivot test set. We reconstruct a fully aligned English-supervised dataset from PHOENIX-2014-T release-v3 annotations (train 7,096 / dev 519 / test 642), preserving pre-extracted visual features. Engineering contributions include checkpoint-safe resumable training and robust annotation-alignment logic for heterogeneous formats. The validated English outputs serve as pivot text for a cascaded multilingual extension in which a post-SLT text translation model maps SLT outputs to additional target languages without retraining the visual backbone. A reproducible end-to-end workflow is provided.