Systematic Exploration of Computational Music Structure Research

In this work we present a framework containing open source implementations of multiple music structural segmentation algorithms and employ it to explore the hyper parameters of features, algorithms, evaluation metrics, datasets, and annotations of this MIR task. Besides testing and discussing the relative importance of the moving parts of the computational music structure eco-system, we also shed light on its current major challenges. Additionally, a new dataset containing multiple structural annotations for tracks that are particularly ambiguous to analyze is introduced, and used to quantify the impact of specific annotators when assessing automatic approaches to this task. Results suggest that more than one annotation per track is necessary to fully address the problem of ambiguity in music structure research.