Emotion Sentiment Analysis in Turkish Music

Abstract

Music has become an indispensable element of human life. Its rhythms, melodies, and harmonies resonate deeply within us, touching our emotions and echoing sentiments in us. In recent years, many music emotion sentiment classifications in different languages have been collected; however, Turkish music has not been explored in depth to our knowledge. We focus on exploring machine learning algorithms to classify Turkish music audio into distinct mood categories. We use a Kaggle dataset, that contains verbal and non-verbal music from different genres of Turkish music. The four classes of emotion in our study include happy, sad, angry, and relaxed. The dataset contains 400 samples with a duration of 30 seconds for each sample. It consists of 50 different acoustic features such as Mel Frequency Cepstral Coefficients, Tempo, Chromagram, Spectral, and Harmonic features. XGBoost, CatBoost, Gradient Boosting, Random Forest, Decision Tree, and Gaussian Naïve Bayes machine learners are used for training and testing the data. Our empirical results show that the CatBoost model has the best performance overall with the area under the ROC curve score of 0.95 and the accuracy, precision, recall, and F1-score of about 80%. Future work includes investigating other emotion-related datasets and other machine learners.