Pengembangan Platform Web Berbasis Large Language Models (LLMs) Terintegrasi Refreshable Braille Display untuk Pengguna Tunanetra

Dameswari, Amelia Rinjani (2026) Pengembangan Platform Web Berbasis Large Language Models (LLMs) Terintegrasi Refreshable Braille Display untuk Pengguna Tunanetra. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5023221027-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Aksesibilitas dokumen digital bermateri matematika bagi penyandang tunanetra masih menjadi tantangan besar akibat keterbatasan pembaca layar dalam menginterpretasikan notasi matematika non-linear dan elemen visual seperti grafik. Penelitian ini bertujuan untuk merancang dan mengimplementasikan VOXORA, sebuah platform web berbasis Large Language Models (LLM) yang melakukan remediasi dokumen secara otomatis dan terintegrasi dengan perangkat Refreshable Braille Display (RBD) 5 sel melalui ekosistem EduBraille. Sistem dikembangkan dengan arsitektur dual-pipeline, menggunakan jalur DOCX berbasis teks XML untuk akurasi simbol dan jalur PDF berbasis gambar visual 150 DPI untuk pemrosesan tata letak halaman utuh. Modul Tanya Dokumen berbasis Retrieval-Augmented Generation (RAG) dan Voice User Interface (VUI) diintegrasikan dengan penundaan suara 1,5 detik guna mencegah bentrokan audio dengan pembaca layar NVDA. Evaluasi kualitas remediasi pada 40 sampel dokumen STEM menggunakan konfigurasi prompt final dua tahap (P4) mencapai nilai Missing Equation Rate (MER) 0,00%, Symbol Preservation Rate (SPR) gabungan 99,13%, dan Hallucination Rate (HR) 0,00%. Dari sisi teknis, skor rata-rata Lighthouse Accessibility Audit mencapai 94,83 dari 100. Pengujian penerimaan pengguna (User Acceptance Testing/UAT) terhadap tujuh responden tunanetra menunjukkan peningkatan skor System Usability Scale (SUS) dari 71,43 (Good) pada laptop menjadi 89,50 (Excellent) pada ponsel (HP), sejalan dengan familiaritas layar sentuh dan screen reader seluler (TalkBack dan Jieshuo). Skor aksesibilitas subjektif mencapai 4,17 dari 5,00, dan Task Completion Rate (TCR) mencapai 88,57% dengan 100% penyelesaian skenario tugas. Deskripsi visual terbukti efektif sebagai jembatan informasi pertama (primary information bridge). Secara keseluruhan, VOXORA terbukti layak dan efektif sebagai prototipe platform aksesibilitas dokumen matematika yang inklusif.
=====================================================================================================================================
Digital accessibility of mathematics documents for visually impaired individuals remains challenging due to screen reader limitations in interpreting non-linear notations and visual elements. This study designs and implements VOXORA, a web-based platform utilizing Large Language Models (LLMs) for automated document remediation, integrated with a 5-cell Refreshable Braille Display (RBD) via the EduBraille ecosystem. The system features a dual-pipeline architecture (XML text-based DOCX and visual image-based PDF) along with a Retrieval-Augmented Generation (RAG) chat module and a Voice User Interface (VUI) with a 1.5-second speech delay to prevent audio collisions with the NVDA screen reader. Quality evaluation of document remediation across 40 STEM samples using the final multi-stage prompt (P4) demonstrated high reliability, achieving a Missing Equation Rate (MER) of 0.00%, a combined Symbol Preservation Rate (SPR) of 99.13%, and a Hallucination Rate (HR) of 0.00%. In terms of technical web accessibility, VOXORA achieved an average Lighthouse accessibility audit score of 94.83 out of 100 (Very Good category). User Acceptance Testing (UAT) with seven visually impaired respondents (including a Mathematics teacher, an English teacher, students, and a university student) revealed a significant increase in System Usability Scale (SUS) scores, jumping from 71.43 (Good category) on laptops to 89.50 (Excellent category) on smartphones (HP), aligning with mobile touch screen familiarity. The subjective accessibility rating achieved an average score of 4.17 out of 5.00, while the Task Completion Rate (TCR) reached 88.57% with a 100% overall task completion rate. Qualitative analysis proved that image descriptions serve effectively as a primary information bridge and a collaborative learning tool. Overall, the VOXORA platform is proven to be an effective prototype for inclusive mathematics document accessibility.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Aksesibilitas Digital, Large Language Models, Refreshable Braille Display, Teknologi Asistif, Tunanetra, Assistive Technology, Digital Accessibility, Large Language Models, Refreshable Braille Display, Visual Impairment
Subjects:	T Technology > T Technology (General) > T58.5 Information technology. IT--Auditing T Technology > T Technology (General) > T59.7 Human-machine systems.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Biomedical Engineering > 11410-(S1) Undergraduate Thesis
Depositing User:	Amelia Rinjani Dameswari
Date Deposited:	31 Jul 2026 03:00
Last Modified:	31 Jul 2026 03:00
URI:	http://repository.its.ac.id/id/eprint/140336

Actions (login required)

View Item