Local Group Delay Based Vibrato and Tremolo Suppression for Onset Detection

In this paper we present a new vibrato and tremolo suppression technique for onset detection. It weights the differences of the magnitude spectrogram used for the calculation of the spectral flux onset detection function on the basis of the local group delay information. With this weighting technique applied, the onset detection function is able to reliably distinguish between genuine onsets and spectral energy peaks originating from vibrato or tremolo present in the signal and lowers the number of false positive detections considerably. Especially in cases of music with numerous vibratos and tremolos (e.g. opera singing or string performances) the number of false positive detections can be reduced by up to 50% without missing any additional events. Performance is evaluated and compared to current state-of-the-art algorithms using three different datasets comprising mixed audio material (25,927 onsets), violin recordings (7,677 onsets) and solo voice recordings of operas (1,448 onsets).