Abstract:

Generative Artificial Intelligence (AI) has significantly lowered the barriers to creating personalized educational software. In our previous research, we introduced the Complete Artifact Workflow (CAW), demonstrating how non-technical stakeholders (parents, educators) can independently generate standalone, offline web applications for children with special educational needs. CAW democratizes the development process for static educational tools, such as arithmetic or calendar learning applications, offering a fast, accessible, and cost-free alternative to traditional software engineering. However, extending this approach to more advanced, voice-interactive assistive tools reveals architectural limitations.

Children with special educational needs often participate in adapted or inclusive educational programs that require customized tools supporting literacy and fine motor skill development. This paper presents a design-oriented case study in which a prototype voice-interactive assistive learning application was developed via CAW. The application supports reading and typing practice through bidirectional audio interaction: Text-to-Speech (TTS) dictates words that the child types using a custom virtual keyboard, while Speech-to-Text (STT) listens to and validates the child’s reading aloud. Although the application was initially implemented as a purely client-side web application following the CAW approach, an architectural analysis revealed several critical technical and practical constraints.

First, achieving strict offline execution—a core CAW requirement intended to preserve the privacy of children’s voice data—is fundamentally hindered by the current implementation of the Web Speech API. Modern browsers route STT processing to external cloud services, effectively operating as a “black box” that prevents true local execution. Second, reliance on operating system–level TTS engines introduces user experience fragmentation. Client-side web applications cannot programmatically ensure consistent voice parameters, such as a specific warm female voice, which is a vital pedagogical factor for children. Finally, while the Web Speech API is highly accessible and free to use, transitioning to native desktop or mobile applications to address privacy and usability limitations introduces new challenges. Offline speech processing requires integrating third-party acoustic models, which raises issues related to software licensing, increased development complexity, and potential costs that contradict the accessibility goals of the CAW approach.

The findings highlight a structural boundary for CAW. While it successfully enables independent end-user generation of fully local educational web applications, extending this paradigm to voice-interactive assistive systems introduces architectural dependencies that cannot be resolved within the constraints of standard browser environments. These results align with emerging discussions on local-first AI in education, which emphasize the importance of processing sensitive student data directly on the learner’s device. Future research should therefore investigate the feasibility of adapting the CAW methodology to support the automated generation of native applications capable of integrating fully local speech processing while preserving the accessibility and affordability of the original workflow.

Keywords:

Assistive Learning, Special Educational Needs, Complete Artifact Workflow, AI in Education, AI Chatbot, Generative AI.