Perbandingan metode voting feature intervals dengan jaringan saraf tiruan dalam mengklasifikasi genre musik

Abstract

Music genre is one of the important descriptions that have been used to classify digital music. The aim of this research is to compare Voting Feature Intervals (VFI) methode with the Neural Network (NN) methode in classifying music genre. There are 12 scenarios of feature extractions in this research. Three variations of MFCC coefficient number (7, 13 and 20 coefficients) and four variations of music length (1, 5, 10, and 30 seconds). From each of the feature vector, mean was calculated. For the NN methode after the feature vectors were extracted, normalization was applied using the cumulative normal distribution methode. This research shown that the optimal number of MFCC coefficients was 13 coefficients. NN predictions were better than VFI predictions. NN has an accuracy up to 95% which was obtained by using 30 neurons of hidden layer, 10 seconds length of music and 13 MFCC coefficients. While the VFI has an accuracy up to 85% which was obtained by using 30 seconds length of music and 7 MFCC coefficients. Both experiments that used 13 and 20 coefficients of MFCC feature obtained same accuracy using the NN method. Classic genre has an accuracy of 100% in VFI. The reliability of the system was 57,14% for disco up to 94,44% for classic.