Voice Driven UML Modeling System for Visually Impaired Students in Software Engineering Education

Visual impairment limits students' access to object-oriented modeling environments that rely on graphical interfaces. Advances in Text-to-Speech (TTS) and Speech-to-Text (STT) technologies make it possible to consider mechanisms to compensate for this barrier, yet their adoption in educational platforms lacks evidence-based guidelines. This work aims to design, implement, and validate a voice-controlled UML modeling system that enables visually impaired students to create, edit, and query class and use-case diagrams. A four-phase methodology was followed: (i) requirements elicitation; (ii) modular architectural design with grammatical validation and the use of large language models (LLMs); (iii) implementation in Python 3.11, FastAPI, and cloud-based STT/TTS services; and (iv) technical evaluation using recognition accuracy, latency, and task-time metrics. The prototype executed 20 critical commands with a recognition accuracy of 97% and maintained full syntactic coherence in the UML models generated. The average times to complete creation, editing, and navigation tasks were 4.23 s, 6.78 s, and 5.24 s, respectively. The average TTS latency (1468 ms) exceeded the 500 ms target, identifying the NLP module as the main bottleneck. The system demonstrates technical feasibility and adheres to the defined accessibility guidelines (WCAG 2.2). Future improvements will focus on reducing TTS latency, expanding the command repertoire, and conducting large-scale usability evaluations (SUS).