A stream segregation algorithm for polyphonic music databases

Most of the existing algorithms for music information retrieval are based on string matching. However, some searching results are perceptually insignificant in the sense that they cannot really be heard, owing to negligence of how people perceive music. When listening to music, it is perceived in groupings of musical notes called streams. Stream-crossing musical patterns are perceptually insignificant and should be pruned out from the final results. Stream segregation should be added as a pre-processing or post-processing step in existing retrieval systems in order to improve the quality of retrieval results. The key ideas are: (a) representation of music in the form of events, (b) formulation of the inter-event and the intercluster distance functions based on the findings in auditory psychology, and (c) application of the distance functions in the adapted single-link clustering algorithm without input of number of clusters. Experiments are performed on real music data to verify our proposed method.