Implementation and Performance Analysis of Mel Frequency Cepstral Coefficient Features in Dangdut Music Sub-Genre Classification

Abstract

This study investigates the classification of dangdut music sub-genres using Mel-Frequency Cepstral Coefficients (MFCC) and machine learning approaches. The objective is to evaluate the effectiveness of MFCC in representing audio characteristics and to compare the performance of several classification algorithms, including K-Nearest Neighbor (K-NN), hybrid K-NN optimized with Genetic Algorithm (GA), Support Vector Machine (SVM), and Decision Tree. The dataset consists of 730 audio samples with a duration of 30 seconds each, categorized into three sub-genres: classic dangdut, rock dangdut, and koplo dangdut. The research process includes audio segmentation, extraction of 13 MFCC coefficients, data normalization, train-test splitting (70:30), and performance evaluation using accuracy, precision, recall, F1-score, confusion matrix, and cross-validation. The results indicate that MFCC provides discriminative feature representations, as demonstrated by improved cluster separation in PCA after normalization. Among the evaluated models, hybrid K-NN with GA achieved the highest accuracy of 98.90%, outperforming conventional K-NN, SVM, and Decision Tree. Confusion matrix analysis showed that most samples were correctly classified, with only minor misclassifications between sub-genres sharing similar audio characteristics. Furthermore, consistently high cross-validation accuracies and low standard deviation values confirmed good generalization capability and suggested the absence of significant overfitting. Overall, the findings demonstrate that MFCC is an effective feature for dangdut music sub-genre classification, while normalization and GA-based optimization significantly improve classification performance and model robustness.