Melodic Segmentation: Evaluating the Performance of Algorithms and Musical Experts Belinda Thom, Christian Spevak, Karin Höthker Institut für Logik, Komplexität und Deduktionssystene, Universität Karlsruhe Outline Segmentation Algorithms Segmentation Data Grouper LBDM Other Approaches Essen Folk Song Collection Musician Corpus Experiments Algorithm - Grouper Designed for monophonic music Based on three rules: Gap Rule Phrase Length Rule Metrical Parallelism Rule Calculates a gap score for pairs of notes Assigns a penalty to boundaries different from an ideal length Combines local view with higher level metric structure Can handle MIDI data Algorithm - LBDM Assigns boundary strength to pairs of notes Quantifies how discontinuous each note pair is Based on two rules: Change Rule Proximity Rule Boundary strength calculations provide insight Can handle MIDI data Algorithm – Other Approaches Memory-based approach Consideration of harmonic structure Rule-based system Based on multi-layer neural network Segmentation Data Essen Folk Song Collection European Folk Songs EsAC format – recording meter and key, pitches and durations Includes phrase boundaries Musician Corpus 10 musical excerpts chosen 19 musicians identified boundaries and phrase and sub-phrase levels Experiments How Ambiguous is the Segmentation Task? Level of ambiguity varies significantly based on music. How Flexible is an Algorithm? Grouper’s results agree more with the musicians segmentations. Yet, LBDM performs better in certain cases. How Stable is an Algorithm? Decrease in performance with unexpected changes, especially in sub-phrases. Grouper performs slightly better.