Decoding Cultural Music Classification with Machine Learning and Segment Length Analysis

Music is a sound composed with rhythm, melody, or harmony, and it is evolving in time and with the culture of the society. The manual way of classifying and searching of the Ethiopian music is time consuming, and expensive. In this study, an Ethiopian cultural music classification models are proposed to simplify this task. A large number of Ethiopian cultural music are collected from YouTube and other online music database. This data set has 10 classes based on the music genre. Eight machine learning algorithms are employed to build the classification models: Logistic Regression, Naive Bayes, KNN, MLP, SVM, Decision Trees, Random Forest, and Adaboost algorithms. The models are optimized using manual hyperparameter tuning, randomized search, grid search, genetic algorithm (TPOT classifier), Bayesian (hyperopt), and Optuna optimization techniques. Based on the experiments, Random Forest outperforms the other algorithms with 80% accuracy. Statistical analysis using one-way ANOVA confirmed that optimization significantly improved classification performance (p < 0.05). Confusion matrix analysis revealed that certain regional styles, such as Gojam and Somali, were more prone to misclassification due to overlapping rhythmic and tonal features. The results demonstrate that machine learning can effectively classify Ethiopian cultural music and provide a foundation for developing intelligent music retrieval and recommendation systems.