Reliable onset detection scheme for singing voices based on enhanced difference filtering and combined features

Onset detection for singing voices is an important but difficult problem for note segmentation in music search or music transcription. It is primarily because the spectral envelops as well as the shapes of vocal tracks vary frequently so as to produce fake peaks in the detection function that degrade the precision rate for onset detection. This paper proposes an onset detection scheme which utilizes the enhanced difference filter to improve the detection function by accentuating the peaks at the uprising margins of energy, and the adjustment approach for onset locations to compensate the peak shifts. A binary classifier based on GMM is further used to combine relevant features of adjacent peaks so as to make more reliable final decisions. The proposed scheme can improve the detection performance significantly, and achieve 78.9% of precision rate and 76% of recall rate at 77.5% of F-measure on the QBSH database.