Article: A study on piano performance training and classroom feedback methods in AI-assisted music performance instruction Journal: International Journal of Reasoning-based Intelligent Systems (IJRIS) 2026 Vol.18 No.28 pp.87 - 99 Abstract: AI-assisted piano instruction requires reliable identification of performance errors and timely conversion of diagnostic evidence into actionable classroom feedback. An event-level multimodal framework is developed using synchronised audio, MIDI, hand-pose, and score information. Its methodological core is a score-conditioned event reliability fusion mechanism, which dynamically weights modality-specific evidence according to score-referenced deviations and branch uncertainty. The model-development dataset contains 156 students from four institutions, with participant-disjoint training and testing. The proposed model achieves an F1-score of 92.2% with an average inference latency of 41.3 ms, while external-institution evaluation retains an F1-score of 90.5%. In the eight-week classroom experiment, rhythm deviation decreases from 79.1 ms to 42.7 ms, note omission rate decreases from 6.9% to 3.1%, and instructor score increases to 86.9 points. The framework is intended for technical performance diagnosis and formative feedback, covering pitch, timing, fingering, dynamics, pedal use, legato continuity, while higher-level artistic interpretation remains under instructor judgment. Inderscience Publishers - linking academia, business and industry through research

Title: A study on piano performance training and classroom feedback methods in AI-assisted music performance instruction

Authors: Hua Wang

Addresses: School of Music, Nangjing University of the Arts, Nanjing, 210013, Jiangsu, China

Abstract: AI-assisted piano instruction requires reliable identification of performance errors and timely conversion of diagnostic evidence into actionable classroom feedback. An event-level multimodal framework is developed using synchronised audio, MIDI, hand-pose, and score information. Its methodological core is a score-conditioned event reliability fusion mechanism, which dynamically weights modality-specific evidence according to score-referenced deviations and branch uncertainty. The model-development dataset contains 156 students from four institutions, with participant-disjoint training and testing. The proposed model achieves an F1-score of 92.2% with an average inference latency of 41.3 ms, while external-institution evaluation retains an F1-score of 90.5%. In the eight-week classroom experiment, rhythm deviation decreases from 79.1 ms to 42.7 ms, note omission rate decreases from 6.9% to 3.1%, and instructor score increases to 86.9 points. The framework is intended for technical performance diagnosis and formative feedback, covering pitch, timing, fingering, dynamics, pedal use, legato continuity, while higher-level artistic interpretation remains under instructor judgment.

Keywords: AI-assisted instruction; piano performance training; multimodal fusion; error detection.

DOI: 10.1504/IJRIS.2026.157152

International Journal of Reasoning-based Intelligent Systems, 2026 Vol.18 No.28, pp.87 - 99

Received: 30 Jun 2026
Accepted: 13 Aug 2026
Published online: 20 Oct 2026 *

Title: A study on piano performance training and classroom feedback methods in AI-assisted music performance instruction

Keep up-to-date