Genre Classification using Feature Extraction and Deep Learning Techniques

Abstract

Music genre refers to categorisation of music on the basis of interaction between artists, market forces, and culture. It helps to organize music into collections by indicating similarities between compositions or musicians. Automatic genre classification is non-trivial as it is difficult to distinguish between different genres. Many times the boundaries are not clearly defined and genres are overlapping. In this paper we present a novel approach to classify a list of songs present on Spotify into mainly four genres-Christian, Metal, Country, Rap. Two different kinds of data - lyrics and album artwork are used for the classification process. Two Natural Language Processing techniques namely bag-Of-Words and Term Frequency-Inverse Document Frequency (TFIDF) are used to process the lyrical data. We apply machine learning algorithms like Random Forest, Support Vector Machine (SVM), Naive Bayes, Linear Support Vector Classifier (Linear SVC) and eXtreme Gradient Boosting (XGBoost) on lyrical data and Deep Convolutional Neural Network (CNN) on the album artwork to predict the genre. On application of machine learning algorithms on lyrical data obtained from bag-Of-Words a mean precision of 75.96% and a mean f-score of 75.92% is achieved. On application of the same set of algorithms on lyrical data obtained from TFIDF a mean precision of 76.85% and a mean f-score of 77.38% is achieved. In both cases XGBoost outperforms all the other algorithms giving a maximum precision of 79.30% and 80.16% and a maximum f-score of 79.6% and 84.09% for bag-Of-Words and TFIDF respectively. On application of deep neural network on album artwork, a precision of 82.46% and a f-score of 81.84% is achieved.