An Adaptive Human-in-the-Loop Generative AI Framework for Professional-Supported Pronunciation Exercise Design: Ex Ante Evaluation in Down Syndrome

Abstract

Personalized pronunciation-exercise design requires flexible content, professional oversight, and mechanisms that can adapt recommendations without transferring therapeutic authority to artificial intelligence. This study presents Speak4All, a cloud-deployed human-in-the-loop generative AI framework designed to support professionals in preparing, personalizing, reviewing, and adapting pronunciation-oriented exercises. Although the platform can support pronunciation activities for primary-school children with different speech-related needs, the present evaluation focuses on contexts involving children with Down syndrome. Following a Design Science Research approach, the framework integrates structured profiles, professionally defined targets, longitudinal assessment history, interpretable difficulty recommendations, deterministic generation constraints, and explicit Accept/Modify/Reject control points. The study follows an ex ante evaluation strategy conducted before involving children as research participants. Two complementary scenarios were considered: an exploratory eight-participant stakeholder evaluation conducted at a volunteer center in Ecuador and a remote evaluation of the cloud-deployed system with 25 anonymous volunteers with relevant professional backgrounds, including special-education teachers working with children with Down syndrome. In the controlled evaluation, the adaptive framework achieved 84.8% acceptance without editing, 100.0% post-validation hard-constraint compliance, and 87.0% exact agreement between adaptive recommendations and independent professional judgments, with a weighted agreement coefficient of κw=0.848. The stakeholder evaluation provided complementary evidence of perceived personalization, usefulness, satisfaction, progress monitoring, and multimedia support. These findings support the technical feasibility and professional acceptability of the framework as a decision-support environment but do not establish clinical effectiveness or improvements in children’s speech. Future prospective studies should evaluate supervised use with children using validated speech-related outcome measures.