Selective Acquisition Techniques for Enculturation-Based Melodic Phrase Segmentation

Automatic melody segmentation is an important yet unsolved problem in Music Information Retrieval. Research in the field of Music Cognition suggests that previous listening experience plays a considerable role in the perception of melodic segment structure. At present automatic melody segmenters that model listening experience commonly do so using unsupervised statistical learning with ‘non-selective’ information acquisition techniques, i.e. the learners gather and store information indiscriminately into memory. In this paper we investigate techniques for ‘selective’ information acquisition, i.e. our learning model uses a goaloriented approach to select what to store in memory. We test the usefulness of the segmentations produced using selective acquisition learning in a melody classification experiment involving melodies of different cultures. Our results show that the segments produced by our selective learner segmenters substantially improve classification accuracy when compared to segments produced by a nonselective learner segmenter, two local segmentation methods, and two naı̈ve baselines.