Application of Music Industry Based on the Deep Neural Network

Abstract

After entering the digital era, digital music technology has prompted the rise of Internet companies. In the process, it seems that Internet music has made some breakthroughs in business models; yet essentially, it has not changed the way music content reaches users. In the past, different traditional and shallow machine learning techniques are used to extract features from musical signals and classify them. Such techniques were cost-effective and time-consuming. In this study, we use a novel deep convolutional neural network (CNN) to extract multiple features from music signals and classify them. First, the harmonic/percussive sound separation (HPSS) algorithm is used to separate the original music signal spectrogram into temporal and frequency components, and the original spectrogram is used as the input of the CNN. Finally, the network structure of the CNN is designed, and the effect of different parameters on the recognition rate is investigated. It will fundamentally change the way music content reaches music users and is a disruptive technology application for the industry. Experimental results show that the proposed recognition rate of the GTZAN dataset is about 73% with no data expansion.