Music signal recognition aids based on convolutional neural networks in music education

With the growth of diverse music information processing needs, music signal recognition technology has become more and more important in music education and music industry. In this study, a music signal recognition aider using convolutional neural network is proposed, and firstly, the logarithmic frequency domain filter bank and double-layer ReLU network are used to extract the pitch features in the music signal. Subsequently, the benchmark convolutional neural network model is constructed, and the constant Q transform is used to process the obtained features to generate a harmonic sequence matrix. Finally, a two-level classification model strategy is used to improve instrument signal recognition. In terms of pitch feature extraction, the accuracy of the logarithmic frequency domain filter group was 74.59 % and 77.03 % respectively under the frame length of 2048 and 8192, which was more effective than the double-layer ReLU network. Experimental results based on different harmonic mapping matrix levels showed that these harmonic mapping matrices had a significant impact on the recall and accuracy of different musical instruments, such as the F1 score of 0.936 for pianos. In the verification of the two-level classification model, the overall accuracy was improved from 0.848 to 0.880 of the benchmark model, which proved the effective improvement of multi-instrument music signal generalization recognition. The research contribution is to improve the ability of pitch feature extraction and establish a more efficient classification model for multi-instrument music signals. These contributions fill the research gap in extracting the pitch and part information of multiple instruments quickly and accurately in complex music works, provide powerful technical support for music analysis and understanding in music education, and innovatively promote the development of music information retrieval technology.