Beat tracking using group delay based onset detection

This paper introduces a novel approach to estimate onsets in musical signals based on the phase spectrum and specifically using the average of the group delay function. A frame-by-frame analysis of a music signal provides the evolution of group delay over time, referred to as phase slope function. Onsets are then detected simply by locating the positive zero-crossings of the phase slope function. The proposed approach is compared to an amplitude-based onset detection approach in the framework of a state-of-the-art system for beat tracking. On a data set of music with less percussive content, the beat tracking accuracy achieved by the system is improved by 82% when the suggested phase-based onset detection approach is used instead of the amplitudebased approach, while on a set of music with stronger percussive characteristics both onset detection approaches provide comparable results of accuracy.QC 20161031