MUSIC GENRE CLASSIFICATION USING ML

Abstract

Music genre classification is well known problem in Music Information Retrieval systems. It has various applications which start from autonomous tagging of unknown music files to provide best recommendations in several applications. Here in this project, we solved the music genre classification problem using different machine learning algorithms such as k-nearest neighbours (k-NN), Support Vector Machine (SVM), Random Forest and XGboost.For experimental purpose we used GTZAN dataset. In this paper the featureshas been extractedwhich includes MFCC, Chroma,Spectral Centroid, Mel spectrogram and Spectral Contrast from GTZAN dataset. The extracted features are high dimensionaland traditional dimension reduction techniques are inefficient on these features, so we solved the dimensionality reduction problem using statistical operations namely mean, median,maximum, mean absolute deviation, standard error of mean, standard deviation,cumulative sum and cumulative maximum on features. This solution not only reduces the dimensions but also keeps the valuable information intact. In this paper, we have also presented the comparative analysis on different combinations of features, by applying various algorithms of machine learning. After performing various experiments, it was found that SVM with Poly kernel has given the best results. The overall improvement in classification that we got is an average accuracy of 80.25% and maximum accuracy of 91% on all 10 musicgenre.