Real-Time Assessment of Piano Sight-Reading Ability in College Students Using CRNN

This study proposes a real-time piano sight-reading assessment framework based on a multi-task convolutional recurrent neural network (MT-CRNN) for objective and intelligent evaluation of college students’ performance ability. The proposed architecture integrates a shared convolutional encoder with parallel decoding branches to simultaneously perform note transcription and musical expression parameter regression. A music rule-based feature fusion module is further designed to align discrete note events with continuous acoustic features, generating a unified representation for multidimensional performance analysis. Based on this representation, an evaluation system is constructed to assess pitch accuracy, rhythmic stability, dynamic control, phrase coherence, and overall musicality. Experimental validation was conducted on a dedicated piano sight-reading dataset containing 400 performance recordings from undergraduate music students. Results demonstrate that the proposed method achieves pitch and duration F1-scores of 95.3% and 94.1 %, respectively, outperforming conventional transcription-based approaches. The generated evaluation scores exhibit strong agreement with expert assessments, with Spearman correlation coefficients reaching 0.84 for dynamic control, 0. 82 for phrase coherence, and 0.79 for overall musicality. Furthermore, the system processes an average 46.8-second performance in only 2.75 seconds, satisfying real-time feedback requirements. The proposed framework provides an effective engineering solution for automated music performance assessment by combining deep learning, acoustic signal analysis, and knowledge-guided evaluation mechanisms, offering practical support for intelligent music education systems.