Research Of Music Genre Classification Based On Deep Neural Network

Nowadays,with the rapid development of Internet,online music service has become the most convenient and important way for people to listen to music.Given to the large amount of online music,the performance of music information retrieval system is critical to the quality of music service.As an important task of music information retrieval,music genre classification has become a hot research topic in recent years.Music signal has a complex frequency structures and rich semantic information,thus it's important to find salient features from music for the automatic genre classification task.In this paper,deep learning methods are used to model music signal.Our experiments showed that it is possible to learn salient features from low-level spectrum of music automatically by using the proposed deep learning models.Specifically,innovations of this work include:1.The semantic information extraction is studied,the algorithm based on recurrent neural networks is proposed.Traditional audio features extraction is usually based on the assumption of the short-time stability of audio signal and thus some of the semantic information which is stored in the context of a relatively long period of time is missed.Making use of the memory characteristics of recurrent neural networks to learn and obtain the context information from a sequence of the short time features may help improve the classification accuracy.The experimental results show that the semantic features obtained by the recurrent neural networks have better representation of the characteristics of music piece from an analysis window,compared with the statistic of the original features.The classification accuracy reached 81.85% and 83.70% in GTZAN and ISMIR2004 datasets respectively.2.Convolutional neural networks and a number of other techniques are proposed to further improve the system accuracy.The spectrum descripts the time-frequency characteristic of audio signal,while convolution neural networks can learn local features about frequency distribution and variety from spectrum.In addition,the high-level representations can be learned from the combination of different local features by stacked convolutional layers.Residual units and stochastic depth strategy are proposed in the training of convolution networks to further improve the performance.Finally,an end to end algorithm,the convolutional recurrent neural network is proposed by combining the convolution and recurrent network.In the experiment,the convolutional recurrent neural network has the best performance among all the network structures in this paper.The classification accuracy in GTZAN and ISMIR2004 reached 88.16% and 89.93% respectively,better than the existing classification algorithm based on convolution network and handcrafted features.This thesis exploits novel algorithm for music genre classification,which would be helpful for the research of high level semantic features extraction,and promote the employment and exploration of deep learning theory in music signal analysis.