Cognition-based segmentation for music information retrieval systems

Background in computer science. This paper investigates the generic problem of model selection in the specific context of Music Information Retrieval (MIR). In MIR research, similarity measures are developed for ranking musical items with respect to their relevance to a user’s musical query. The application of such similarity measures in MIR systems typically requires musical works to be divided into more manageable units. This involves two tasks: melody segmentation and voice separation. For both of these tasks, several solutions have been proposed in the symbolic domain. It seems reasonable to assume that those solutions that are most in accordance with human performance will result in the best ranking of retrieval output. As a first step towards this goal, this paper describes the evaluation of ten prominent methods against human performance. Background in cognition. Human listeners generally possess two functions that allow them to process a continuous stream of music into understandable units: the ability to perceive multiple, successive tones as one coherent melodic phrase (melody segmentation) and the ability to differentiate melody notes from harmony notes (voice separation). Algorithms for mimicking these human functions have been developed from two perspectives: model-driven, taking Gestalt principles as a starting-point; and data-driven, inferring rules by learning from large amounts of data. One method excepted, this research focuses o