A Comparison of Score-Level Fusion Rules for Onset Detection in Music Signals

Finding automatically the starting time of audio events is a difficult process. A promising approach for onset detection lies in the combination of multiple algorithms. The goal of this paper is to compare score-level fusion rules that combine signal processing algorithms in a problem of automatic detection of onsets. Previous approaches usually combine detection functions by adding these functions in the time domain. The combination methods explored in this work fuse, at score-level, the peak score information (peak time and onset probability) in order to obtain a better estimate of the probability of having an onset given the probability estimates of multiple experts. Three state-ofthe-art spectral-based onset detection functions are used: a spectral flux detection function, a weighted phase deviation function, and a complex domain detection function. Both untrained and trained fusion rules will be compared using a standard data set of music excerpts.