Comparison of music genre classification using Nearest Centroid Classifier and k-Nearest Neighbours

Abstract

Music genre is getting complex from time to time. As the size of digital media grows along with amount of data, manual search of digital audio files according to its genre is considered impractical and inefficient; therefore a classification mechanism is needed to improve searching. Zero Crossing Rate (ZCR), Average Energy (E) and Silent Ratio (SR) are a few of features that can be extracted from digital audio files to classify its genre. This research is conducted to classify music from digital audio (songs) into 12 genres: Ballad, Blues, Classic, Harmony, Hip Hop, Jazz, Keroncong, Latin, Pop, Electronic, Reggae and Rock using above mentioned features, extracted from WAV audio files. Classification is performed several times using selected 3, 6, 9 and 12 genres respectively. The result shows that classification of 3 music genres (Ballad, Blues, Classic) has the highest accuracy (96.67%), followed by 6 genres (Ballad, Blues, Classic, Harmony, Hip Hop, Jazz) with 70%, and 9 genres (Ballad, Blues, Classic, Harmony, Hip Hop, Jazz, Keroncong, Latin, Pop) with 53.33% accuracy. Classification of all 12 music genres yields the lowest accuracy of 33.33%. The test results with the k-Nearest Neighbours algorithm to 120 songs for k = 3 accuracy reaches 22.5%, k = 5 accuracy reaches 22.5%, k = 7 accuracy reaching 26.7% and k = 9 accuracy reaches 26.7 %. Results showed that genre classification by matching the shortest distance through the centre of the class, yields better results than using the k-NN algorithm.