Abstract
Audio key finding is an integral step in content-based music indexing and retrieval. In this paper, we present a system that combines ensemble learning with an existing model-based key finding algorithm: the Fuzzy Analysis Center of Effect Generator algorithm. We demonstrate the manner in which AdaBoost improves the accuracy of FACEG using a dataset containing 2785 audio excerpts of real performances composed by Bach and Mozart. Two sets of experiments were conducted: intra-system comparison examining the effect of different settings in FACEG/AdaBoost, and inter-system comparison comparing FACEG/AdaBoost with the key finding implementation in Music Information Retrieval (MIR) toolbox. When FACEG is executed to generate keys at multiple stopping points of the excerpt, AdaBoost with multi-views of tonal information improves key detection accuracy up to 35% on the challenging dataset and up to 21% on the entire dataset.