Music Onset Detection Based On Constant Q Transform

Note onset detection is the basic problem of music signal analysis and processing.It is the keyof Content-Based Audio Information Retrieval,especially for Query By Humming. The accuracy ofthe onset detection for each note affect the accuracy of Content-Based Audio Information Retrievalsystem in largely. Note onset detection techniques mostly learn speech end-point detectiontechnology. Firstly, it introduces the research background and significance of note onset detectioncomprehensively, summarizes four steps of this detection system: pre-processing, feature extraction,selection feature equation, peak-picking. Secondly, it introduces several classical algorithms, andanalyzes their advantages and disadvantages. Lastly, it puts forward a more effective,new notedetection algorithm—note onset detection based on Constant Q transform. On the basic of ConstantQ Transform, using the spectral energy and sub-band spectral entropy to detect the note, calculationthe distance of notes, last using hierarchical method to optimize the detection function.In audio signal processing, in order to observe the time position of the high frequencyphenomenon, we should use the narrow time-domain window; In order to understand the lowfrequency phenomenon, we should use the wide time-domain window; Short time Fourier transformalways use the same length of the window,the Constant Q Transform is able to automatically adjustthe window length. This paper is on basic of Constant Q Transform, it‘s spectral energy can reactionthe changes well, but for the―soft‖notes of music signal, it‘s spectral energy don‘t havesignificantly changes, Spectral entropy is related to the randomness of the music signal, it isregardless of the amplitude of the signal, the soft‘note‘s spectral entropy is not necessarily small,just to make up for the drawback of the method of spectral energy. Sub-band spectral entropy is notonly inherited the advantages of spectral entropy, but also have certain resist-noise performance. Inthis paper, the two characteristic parameters of the spectral energy and sub-band spectral entropyare multiplied, the notes onset detection of time is more accurate. Finally, a new optimizationdetection function-hierarchical method, to optimize the detection function, this can make the peaksmore obvious, we can use fixed threshold method to extract local maxima.In the simulation phase, for different types of music signal,using method of the difference ofspectrum based on the short time Fourier transform, spectral energy, the sub-band spectral entropy,integrated of spectral energy and sub-band spectral entropy based on the Constant Q transform tosimulation, analysis the simulation result. The experiment results show that the method proposed in this paper has higher accuracy rate.