Simple Spectrum-Based Onset Detection
Simon Dixon · 2006
In a recent empirical study, various methods for detecting the onset times of musical notes in audio signals were evaluated [1]. The study focussed on published methods based on spectral features such as the magnitude, phase and complex domain representations, and compared existing methods (spectral flux, phase deviation and complex difference) with proposed improvements to these methods (weighted phase deviation, normalised weighted phase deviation and rectified complex difference). Two test sets were used: a set of short excerpts from a range of instruments (1060 onsets), plus a much larger data set of piano music (106054 onsets). Results showed a similarly high level of performance with a magnitude-based (spectral flux), a phase-based (weighted phase deviation) or a complex domain (complex difference) onset detection function. For MIREX 2006, the following five onset detection functions were submitted: spectral flux, complex domain, rectified complex domain, weighted phase deviation and normalised weighted phase deviation.