Can explainable AI scaffolding reduce cognitive load and enhance embodied expression? A quasi-experimental study of...

Exploring how artificial intelligence technology can effectively support music education has become a critical issue in the digital era. This study addresses the high cognitive load and embodied expression difficulties faced by bel canto students when learning multilingual vocal works, designing and validating the educational effectiveness of an explainable AI scaffolding learning system. The study employed a 2 × 3 mixed-factor quasi-experimental design, with 64 college students from the Conservatory of Music allocated to the XAI scaffolding group and the traditional AI-assisted group through stratified assignment for a 16-week intervention on the learning of vocal music works in Italian, German, and French. Data collection integrated multimodal measurements including cognitive load scales, expert ratings, motion capture, and eye-tracking, with analysis using repeated-measures mixed-factor ANOVA, Bootstrap-based mediation analysis, and thematic analysis. The study found that the XAI scaffolding group's cognitive load reduction reached 38.9%, significantly higher than the control group's 17.0% (p < 0.001, d = 2.36, 95% CI [1.72, 2.99]), with extraneous cognitive load reduction of 41.2% being the primary contribution; embodied expression total scores improved by 45.0%, with body posture, facial expression, and gesture use dimensions all significantly superior to the control group; language complexity moderated XAI effects, with the advantage most pronounced in German learning (g = 2.71); the three sequence-disambiguation analyses converged on an independent contribution of language complexity beyond cumulative learning, with the language × group interaction remaining significant after cumulative-exposure adjustment ( ${\eta }_p^2$ = 0.078, p = 0.007), although complete causal separation awaits future counterbalanced replication; cognitive load played a partial mediating role between XAI scaffolding and embodied expression, with indirect effects accounting for 36.6% of total effects; metacognitive level significantly moderated the mediation pathway; qualitative analysis revealed an "understanding-experiencing-integrating" learning pathway mechanism. This study provides empirical evidence of the potential value of explainable AI in vocal education, offering preliminary theoretical insights and practical guidance for transparent design of intelligent music education systems. These findings warrant replication with larger, multi-site samples and counterbalanced sequencing designs before direct generalization to educational practice.