Fusion of signal processing methods with pitch detector for onset detection

Onset detection, the quest for finding “transient” regions in the audio signal, is an important task, which is the base of high-level musical processing tasks. Because the existing algorithms have their own problems respectively, time-domain energy algorithm is efficient in detecting an attack transient event in drum, whereas has a problem in bowed instrument that can be solved by phase-based detection, on the other hand, spectral flux detection which has a poor performance on the percussive instruments is very effective on wind as well as bowed string instruments. In this paper, we linearly combine these methods and add pitch detector to engender an adaptive algorithm. Experiments are conducted on a dataset including four types of instruments. Results show that the F-measure of our algorithm is better than other individual methods.