Note Extraction and Recognition Analysis Based on Music Melody Features

Abstract

This paper extracted and recognized the notes in the music based on the features of the music melody. Firstly, the melodic features and Mel-frequency cepstral (MFCC) features were extracted from the music signal and then combined. A convolutional neural network (CNN) was used as a classifier for note classification and recognition in the music signal. The CNN adopted a structure with three convolutional layers and three pooling layers. The melody features and MFCC features were used to extract convolutional features through convolution kernels in the convolutional layers, followed by compressing these features in the pooling layers. Finally, the note recognition results were outputted in the output layer. Then, simulation experiments were performed using a self-built music library. The performance of the algorithm was tested under different MFCC feature dimensions and CNN activation functions. The algorithm was also compared with the dynamic time warping (DTW) algorithm and the CNN algorithm without music melody features. The results showed that the proposed algorithm had the best performance when the MFCC feature dimension was set to 24, and the CNN activation function was sigmoid; under such conditions, the F-measure was 96.7%.  The performance of the proposed algorithm was the best, regardless of whether it recognized the single-note or multi-note music. The precision for recognizing single notes was 98.3%, the recall rate was 96.5%, and the F-measure was 97.4%. For recognizing multiple notes, the corresponding values were 93.0%, 92.5%, and 92.7%, respectively. However, the performance of the three algorithms was reduced when recognizing multi-note music.