Beatbox Classification to Distinguish User Experiences Using Machine Learning Approaches

Abstract

Research regarding beatbox classification has generated a relatively significant growth in the past decade. Although the differences between contributors’ expertise within a vocal percussion dataset have been mentioned in previous works, the impact of those discrepancies has not been thoroughly investigated. In this study, the authors explore performances of machine learning algorithms for beatbox classification, with an emphasis on prior beatboxing experience affecting dataset. Throughout this study, feature extraction is conducted by the use of 4 methods, i.e. Spectral Centroid, Spectral Magnitude, Spectral Contrast, and MFCC, while machine learning method to perform classification is through the use of KNN (3,5,7), Adaboost, LSVM one-vs-one, LSVM one-vs-rest, SVM one-vs-one, SVM one-vs-rest. This study shows that performing a beatbox classification requires more thought into the differences between the skill level of the dataset (inexperienced and trained/professional). Points of concern include the shorter time span in a trained beatbox dataset to segment and classify before the next onset begins, in which some sounds were even found to be smaller than 0.01 ms. For classification experiments using several feature extraction techniques and machine learning models, experiment results show that MFCC (n_mfcc = 22) delivers the best feature representation for our KNN, multi-class and non-linear SVM classification model.