A Scalable Hybrid Classifier for Music Genre Classification using Machine Learning Concepts and Spark

TitleA Scalable Hybrid Classifier for Music Genre Classification using Machine Learning Concepts and Spark
Publication TypeConference Paper
Year of Publication2018
AuthorsKarunakaran, N.., and A.. Arya
Conference Name2018 International Conference on Intelligent Autonomous Systems, ICoIAS 2018
Date Published2018
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN Number9781538663295 (ISBN)
KeywordsArtificial intelligence, Audio feature extraction, Classification process, Classifiers, Discriminant analysis, Electronic musical instruments, essentia, Extraction, Feature extraction, Fuzzy sets, Information retrieval, K-nearest neighbor classifier, Learning systems, machine learning concepts, Music genre classification, music information retrieval, Naive Bayes classifiers, Nearest neighbor search, Quadratic discriminant analysis

Music genre classification has been a challenging task in the field of Music Information Retrieval (MIR). Using a machine to automate this classification process is a more complex task. Audio feature extraction is the first important step of MIR. Essentia, a tool is used for the audio feature extraction. In this paper, a comparative study of the standard machine learning classifiers including K-Nearest Neighbor Classifier, Support Vector Machine, Naive Bayes Classifier, Neural Network, Quadratic Discriminant Analysis and Fuzzy Classifiers is undertaken in context of music genre classification. In this paper, a two phase hybrid classifier is proposed to overcome the problem of blurry classification of Pop, Rock and Electronic genres. It is observed that these genres are not classified accurately by above mentioned classifiers. The behavior of the proposed classifier on the standard music genre dataset GTZAN Genre Collection (1000 songs) and Free Music Archive (FMA) Dataset (approx. 20K songs) is investigated. The overall result of proposed approach on both the datasets is approximately 90% and 70% respectively. © 2018 IEEE.