Design And Implementation Of A Content-based Music Genre Auto-classification System

With the rapid development of Internet and multimedia technologies,online music services have become one of the most important Internet online services for mass consumers.The scale of the music libraries on the Internet has been extremely large.There are major online music libraries which contain more than millions of tracks.The scale of these music libraries is still rapidly expanding.An important and urgent requirement is to automate the labeling then reasonably organize and classify massive digital music tracks so that consumers can efficiently and quickly retrieve music tracks from massive music libraries according to their personal preferences.In recent years,Music Information Retrieval(MIR)has developed into a remarkable new research field.The content-based music genres auto-classification is an important task of in MIR.This study designed and implemented a content-based music genre auto-classification system.Based on the liquidity over time of music,the author designed a special-structured deep learning classifier,exploiting convolutional neural network and LSTM,using Mel-spectrogram as its input.The effectiveness of this special-structured deep learning classifier was verified on the GTZAN data set;An extra advantage of this deep learning classifier is that it can conveniently process music in real time;Harmonic/percussive separation on Mel-spectrogram and ensemble learning technology were employed in order to further improve the classification accuracy of the deep learning classifier;Sufficient exploration on the prediction performance of the classifier under various classification probability threshold settings was carried out,which gives advices to the classification system to perform music retrieval according to a given target music genre.