Design And Evaluation of Onset Detectors using Different Fusion Policies

ABSTRACTNote onset detection is one of the most investigated tasksin Music Information Retrieval (MIR) and various detec-tion methods have been proposed in previous research. Theprimary aim of this paper is to investigate different fusionpolicies to combine existing onset detectors, thus achiev-ing better results. Existing algorithms are fused using threestrategies, first by combining different algorithms, second,by using the linear combination of detection functions, andthird, by using a late decision fusion approach. Large scaleevaluation was carried out on two published datasets and anew percussion database composed of Chinese traditionalinstrument samples. An exhaustive search through the pa-rameter space was used enabling a systematic analysis ofthe impact of each parameter, as well as reporting the mostgenerally applicable parameter settings for the onset de-tectors and the fusion. We demonstrate improved resultsattributed to both fusion and the optimised parameter set-tings.1. INTRODUCTIONThe automatic detection of onset events is an essential partin many music signal analysis schemes and has various ap-plications in content-based music processing. Different ap-proaches have been investigated for onset detection in re-cent years [1,2]. As the main contribution of this paper, wepresent new onset detectors using different fusion policies,with improved detection rates relying on recent research inthe MIR community. We also investigate different config-urations of onset detection and fusion parameters, aimingto provide a reference for configuring onset detection sys-tems.The focus of ongoing onset detection work is typicallytargeting Western musical instruments. Apart from usingtwo published datasets, a new database is incorporated intoour evaluation, collecting percussion ensembles of Jingju,also denoted as Peking Opera or Beijing Opera, a major