Evaluating Melodic Segmentation

We discuss why evaluating melodic segmentation is difficult when solutions are ambiguous and explore the nature of this ambiguity using a corpus of melodies hand-segmented by musicians. For each melody, the musicians identified different, musically coherent solutions, suggesting that ambiguity should be modeled when assessing how “well” an algorithm segments the same melody. We propose a probabilistic framework for modeling ambiguity that integrates both the segment boundaries and the lengths that the musicians preferred. The framework gives rise to several potential extensions of existing segmentation algorithms.