A Hierarchical Approach to Onset Detection

Onset detection in vocal music and many other instruments is complicated by the possibility of soft transitions between notes. Most systems try to identify onsets within a short-time window as it is easier to define transition functions over a restricted space. However, it may not be possible to detect soft onsets without considering a long-time window, for which defining and computing the transition function can be hard and computationally costly. We present a method which looks for onsets between locations of increasing distance and is able to capture such onsets without considering all the points within the window. For the onset identification function we use both a simple manual function and support vector machines trained using a labelled corpus.