Onset Detection Based on Fusion of Simpls and Superflux

In this paper, an onset detection method based on Simple Partial Least Squares (SIMPLS) is proposed and a system of fusing SIMPLS and SuperFlux is introduced. SIMPLS is an efficient approach to partial least squares regression which has been applied to classification tasks. To detect onsets in an audio file, the file is sampled discretely into frames and the signals are transferred into frequency domain. The information is then fed into SIMPLS to predict a short time frame is onset or not. A score is obtained for each frame as the possibility of being an onset, and post processing is conducted to pick the peaks. SuperFlux is a robust and fast onset detection method proposed by S. Böck et al. We combine the detection results of SIMPLS and SuperFlux on the decision level. Experimental results demonstrate that the proposed method could generate better F1-measure value than the individual methods.