Answers are grounded in this paper and cite their sources.
With increasing use of artificial intelligence for performance evaluation in music education, repeated mistakes may cause frustration, anxiety, and even withdrawal from practice due to lack of proper feedback mechanisms addressing such problems. This study developed a multimodal feedback system based on MIDI performance, facial expression coefficients, and interaction logs where an attention-based BiLSTM predicted valence, arousal, and need for high regulation and then used contextual Thompson sampling to select affect-adaptive feedback. In the pilot study, the dataset consisted of 1,080 practice segments from 45 novice piano players, and the six-week experiment included 90 subjects allocated into affect-adaptive, fixed technical, and metric-only feedback conditions. The full multimodal model obtained the following scores: Macro-F1 = 0.873, AUC = 0.936, ECE = 0.029, and MEEL = 46.7 ms. At posttest, note accuracy for the adaptive condition amounted to 93.1%, while in two other conditions, it amounted to 89.4% and 87.5%. Also, restart latency for the adaptive condition reduced to 7.3 seconds. Moreover, multilevel mediation analysis revealed a significant indirect effect through emotional recovery, β = 0.16, 95% CI [0.07, 0.27]. This research shows that integrating affect recognition with feedback selection is a promising direction for improving computational responsiveness and emotional regulation in music learning. Proposed architecture maintains boundaries set by teacher-reviewed feedback and separates performance and outcome evaluations.
No takes yet. Share an insight, caveat, or question.
Binru Zou (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: