Music Note Onset Detection Based On Sparse Decomposition

Note onset detection is an important issue on content-based music information retrieval. Asnote’s onset is low-level audio feature, note onset detection plays an important role on research ofmusic rhythm analysis, beat tracking and music structure analysis. Note onset detection is also thebasic work of musical topics, such as music emotion classification, packet loss recovery. Focusingon note onset detection, this paper researches in the following aspects:(1) After summarizing the steps of traditional note onset detection: pre-process、signalconversion、feature extraction、generate detection function、peak extraction, this paper uses sparsedecomposition in signal conversion. Sparse decomposition with redundant dictionary can adapt todifferent music signals, seize the inherent characteristics.(2) This paper uses matching pursuit algorithm of sparse decomposition and extracts twofeatures: explanation degree and the energy of coefficient. This paper smoothes detection functionswith Gaussian kernel smooth algorithm. Experiments show that after the processing of Gaussiankernel smoothing, the curve of detection function becomes smoother and peaks become moreprominent.(3) Experiments compare the results of two new algorithms and traditional algorithms,analyze the influences of matching pursuit decomposition times. Experimental results show that thenew onset detection algorithms have higher accuracy and when the decomposition times is small,with the increase of decomposition times, the accuracy of the results improved. However, when thedecomposition times is big enough, with he increase of decomposition times, the accuracy of theresults remains unchanged.