Amplitude modulated sinusoidal modeling for audio onset detection

Onset detection is a key application in music processing. Beat detection algorithms and some music transcribers usually perform onset detection as the starting point of their processing. In music transcription of polyphonic signals, onset detection is very helpful because it aids to place note-event starting times. In this paper, a new technique to implement an onset detection system is proposed. In sinusoidal modelling, the energy burst of non-stationary tones are detected by means of linear prediction in the frequency domain. In frequency, the tone peak and its nearby samples does not match with the window transform when the tone is not stationary at the current frame. This property can be detected whith linear prediction in the frequency domain. When perceptually significant tones are detected as unstable in a time frame, the system alerts about an onset at this frame. The proposed onset detection system is evaluated over two sound databases obtaining encouraging results.