Machine Learning Analysis of Music Based on Music Information Retrieval Tasks

Abstract

Music Information Retrieval (MIR) methods extracts from music high-level information like classification, musical feature extraction, song similarity and tonality. Musical genre is one of the orthodox methods of describing musical content and a significant part of MIR. At present, few MIR research has been done on Nigerian songs. So, this paper proposed to build a genre classification model based on Mel Spectrogram of audio songs. The process first converts ORIN audio dataset to Mel Spectrogram and extract numerical information from it using the Histogram of Oriented Gradient (HOG) and apply machine learning (ML) models to accurately categorize the songs into different genres of Apala, Fuji, Juju, Highlife and Waka. Support Vector Machine (SVM) with 4 different kernels, with 10- cross validation method were applied and assessed based on Accuracy and Receiver operating characteristics (ROC).