부산대학교 도서관

The tonal centre of a piece of music is crucial for perception and understanding, estimating the root note is an important issue in music information retrieval. To determine the root note of a melody, previously published methods generally involve matching the histogram of pitch classes in the melody with pre-calculated pitch class profiles. Performance varies between methods depending on these profiles, as there is no universally agreed-upon set of profiles. Furthermore, these methods do not consider structural information. In this study, we present an ensemble machine learning-based algorithm for root note prediction. In the proposed methodology, we estimate the key of a piece using several existing pitch-class profile algorithms and several simple key-estimation heuristics. We then feed these results into a machine-learning model. Experiments are conducted on the Ceol Rince na hEireann dataset, which contains 1,224 Irish traditional dance music melodies. The dataset is unbalanced, with some possible root notes occurring rarely or never, reflecting common practice in the tradition. To achieve a balanced dataset, we employed both simple oversampling and the Synthetic Minority Over-sampling Technique (SMOTE). The experiments revealed that the Random Forest classifier, trained on a balanced dataset, performed the best. The F1 score was approximately 0.90 when tested on unseen data.