Research On Music Classification Based On Auditory Feature Convolution Neural Network

Since the cassette era,the CD era,and the era of digital music,the amount of music has boomed.Only the classification of massive music resources and the establishment of an efficient music retrieval system can help people quickly search for the desired music.Traditional music classification methods rely on manual annotation.In the era of big data,it is inefficient and unrealistic to classify music through manual annotation.Thus,the automatic classification of music has gradually become a research hotspot and is widely used.Automatic music classification is the basis of fast and effective retrieval of music resources.It has huge potential application requirements,so the automatic music classification method has important research significance.There are two key steps in automatic music classification: feature extraction and classification.In this paper,two key steps of feature extraction and classification are improved based on auditory characteristics.A convolutional neural network classification method based on auditory characteristics is proposed for music classification tasks.The main work and innovation of this article are as follows:(1)In the feature extraction,this paper uses the cochlear filter cepstrum coefficients which is commonly used in speech signal processing to apply it to music feature extraction.Cochlear filter cepstrum coefficient extraction process simulates the process of the human auditory system's perception of sound,making the extracted features more in line with human auditory characteristics.Because the span of the frequency domain of music signal is wider than that of the speech signal,the filter of the low frequency and high frequency is added to the step of extracting the cepstrum coefficient of the cochlea filter,so that the extracted music features are more complete and detailed,which directly improve the quality of feature extraction.(2)In the classification method,the paper draws on the convolutional neural network proposed by Lecun et al.and the local convolution idea proposed by Taigman et al.then proposes and implements a convolutional neural network based on the auditory characteristics for the music classification scene.The traditional convolutional neural network convolution kernel is globally shared,and all frequency domain information is processed in a consistent manner,ignoring the difference in frequency domain information.The convolutional neural network based on auditory characteristics divides the time-frequency features of music into different regions according to the frequency level,and only shares the convolution kernels within a specified region,so that the convolution kernels in different frequency regions learn the required characteristics of their respective frequency regions.(3)In this paper,the music application scene classification experiments are conducted for four playing scenarios in cafes,learning,nightclubs,and sports.The cochlear filter cepstrum coefficients and the convolutional neural network based on auditory characteristics achieve 83.58% accuracy,which is superior to that of the combination of Mel-frequency cepstrum coefficients and traditional convolutional neural networks.