Evaluating the Online Capabilities of Onset Detection Methods

In this paper, we evaluate various onset detection algorithms in terms of their online capabilities. Most methods use some kind of normalization over time, which renders them unusable for online tasks. We modified existing methods to enable online application and evaluated their performance on a large dataset consisting of 27,774 annotated onsets. We focus particularly on the incorporated preprocessing and peak detection methods. We show that, with the right choice of parameters, the maximum achievable performance is in the same range as that of offline algorithms, and that preprocessing can improve the results considerably. Furthermore, we propose a new onset detection method based on the common spectral flux and a new peak-picking method which outperforms traditional methods both online and offline and works with audio signals of various volume levels.