Integrating Segmentation and Similarity in Melodic Analysis

The recognition of melodic structure depends on both the segmentation into structural units, the melodic motifs, and relations of motifs which are mainly determined by similarity. Existing models and studies of segmentation and motivic similarity cover only certain aspects and do not provide a comprehensive or coherent theory. In this paper an Integrated Segmentation and Similarity Model (ISSM) for melodic analysis is introduced. The ISSM yields an interpretation similar to a paradigmatic analysis for a given melody. An interpretation comprises a segmentation, assignments of related motifs and notes, and detailed information on the differences of assigned motifs and notes. The ISSM is based on generating and rating interpretations to find the most adequate one. For this rating a neuro-fuzzy-system is used, which combines knowledge with learning from data. The ISSM is an extension of a system for rhythm analysis. This paper covers the model structure and the features relevant for melodic and motivic analysis. Melodic segmentation and similarity ratings are described and results of a small experiment which show that the ISSM can learn structural interpretations from data and that integrating similarity improves segmentation performance of the model.