In order to solve the problems of strong subjectivity and insufficient personalization in traditional music performance evaluation methods, this study proposes a music teaching evaluation and feedback system based on music signal processing and Convolutional Recursive Hash Network (CRNNH) model. The system first applies wavelet transform to denoise the original performance audio, and uses a pre trained FCN-5 network to extract multi-level feature maps. Then, a double-layer long short-term memory (LSTM) is further introduced to capture the temporal dependence of music signals. The test results of the MagnaTag Tune dataset show that the proposed system achieved a recognition accuracy of 91.4% in 300 recognition tasks, which is 7.2% higher than the system based on deep belief networks (DBN). When the signal-to-noise ratio (SNR) is 20 dB, the proposed system achieves an accuracy of 91.8%, and also maintains a certain advantage when the SNR is -5 dB. In addition, in terms of efficiency, the proposed system only takes 17.0 seconds to complete 300 tasks, while the DBN based system takes 22.8 seconds. In practical music teaching, the proposed system can effectively improve the accuracy of performance error recognition and provide real-time personalized feedback, which helps promote the intelligent and personalized development of music teaching.
