Real-time music score transcription auxiliary teaching system for piano education based on CNN and CQT

Automatic music transcription technology provides key support for intelligent piano education . However , traditional methods have problems such as insufficient transcription accuracy and poor real-time performance. To address these issues, this study designed a method based on the Constant Q Transform ( CQT ) and a lightweight Convolutional Neural Network (CNN). A real-time music score transcription assistance teaching system for piano education based on a CNN (likely a CNN) is proposed. The system first samples piano audio at a 44.1kHz sampling rate, pre-emphasizes it, performs frame- by-frame windowing, and then performs a CQT transform to generate a time-frequency feature map that accurately reflects the piano's pitch characteristics. It then constructs a dual-task lightweight CNN model, separating pitch estimation and onset detection branches after a shared feature extraction network . A joint loss function is used for optimization, keeping the number of parameters below 2M to ensure real-time performance. Experimental results show that on the MAPS dataset, the system achieves a precision of 93.1 %, a recall of 91.9%, and an F1-score of 92.5%, all outperforming other comparable methods . Its average end-to-end latency is only 80ms, with a maximum latency of no more than 90ms. This system combines high transcription accuracy with low latency, providing real-time, accurate assistance for piano instruction and significantly improving the efficiency of piano education.