Music Auto-tagging Based On Generative Adversarial Network

With the advent of the era of informationization and digitization,digital music resources have increased rapidly.In this context,the tags are becoming more and more important as a way of organizing structured information of music.Music annotation refers to the description of music semantics by generating music tags,thereby realizing rapid retrieval,efficient management and personalized recommendation of massive music resources.At present,the methods commonly applied to solve the problem of music tagging include manual tagging and social tagging,but there are cost and quality problems in these methods.Improving the predictive effect of the music auto-tagging model is one of the effective ways to solve this problem.Based on this,this paper proposes music auto-tagging based on generative adversarial network.In this paper,the commonly used Mel cepstrum coefficients are improved for the audio characteristics of music,and the Gamma-chirp Filter Cepstral Coefficient(GCFCC)is introduced to simulate human ear characteristics.For the tag vector of music,the music tag is clustered by the LDA theme model to obtain the topic category.In order to achieve better music auto-tagging effect,this paper applies InfoGAN which is improved based on GAN,and thus finds the mapping relationship between the audio features and semantic features of music.In the experiment,the precision rate,recall rate and F1-measure are used to measure the experimental results.The experimental results obtained by MFCC and GCFCC as audio characteristic parameters are compared and analyzed.At the same time,the experimental results are also compared with other models.The final experimental results show that the improved Mel feature representation and InfoGAN model have achieved good results in the experiment.