Location Constraints for Repetition-Based Segmentation of Melodies

摘要

Repetition-based modelling of melody segmentation relies on identifying find selecting repetitions of melodic fragments. At present, automatic repetition identification results in an overwhelmingly large number of repetitions, requiring the application of constraints for selecting relevant repetitions. This paper proposes constraints based on the locations of repetitions, extending existing approaches on constraints based on repetition length and frequency, and the temporal overlap between repetitions. To test our constraints, we incorporate them in a state-of-the-art repetition-based segmentation model. The original and constraint-extended versions of the model are used to segment 400 (symbolically encoded) folk melodies. Results show the constraint-extended version of the model achieves a statistically significant 14% average improvement over the model's original version.