Audio Acoustic Features Based Tagging and Comparative Analysis of its Classifications

Abstract

Musical genres can be used to distribute and manage music datasets to increase the ease in finding a music piece a person wants to listen to. This paper presents a research for creating a suitable model for genre recognition in audio files using machine learning classifiers on the IRMAS11 https://www.upf.edu/web/mtg/irmas dataset. Python language library pyAudinAnalysls22 https://github.com/tyiannak/pyAudioAnalysis for extracting features from audio files is used. Further, three base classifiers, namely Support Vector Machines (SVM), Decision Tree and Random Forest are also depicted. IRMAS [10] genre dataset provides 6705 audio files of four genres classical, country folk, jazz and pop-rock. Also explored is an ensemble classification model by creating a stack of classifiers for the genre recognition task. Genre classification using SMOTE has been characterized in the confusion matrix. Maximum accuracy of 81.56% using the ensemble classifier is achieved using the proposed methodology.