Research on Neural Network-based Automatic Music Multi-Instrument Classification Approach

Abstract

The automatic classification of multi-instruments plays a crucial role in providing services for music retrieval and recommendation. This paper focuses on automatic multiinstrument classification. Firstly, instrument features were analyzed, and Mel-frequency cepstral coefficient (MFCC) and perceptual linear predictive coefficient (PLPC) were extracted from instrument signals. Features were selected using the entropy weight method. The optimal initial weight threshold of a back-propagation neural network (BPNN) was obtained by utilizing the sparrow search algorithm (SSA), achieving a SSA-BPNN classifier. Experiments were conducted using the IRMAS dataset. The results demonstrated that the combination of MFCC and PLPC selected through the entropy weight method achieved the best performance in automatic multi-instrument classification. The method yielded high P value, recall rate, and F1 value, 0.72, 0.71, and 0.71, respectively. Moreover, it outperformed other algorithms such as support vector machine and XGBoost. These results confirm the reliability of the automatic multi-instrument classification method proposed in this paper, making it suitable for practical applications.