Abstract

Whilst many onset detection algorithms for musical events in audio signals have been proposed, comparative studies of their e#cacy for segmentation tasks are much rarer. This paper follows the lead of Bello et al. 04, using the same hand marked test database as a benchmark for comparison. That previous paper did not include in the comparison a psychoacoustically motivated algorithm originally proposed by Klapuri in 1999, an oversight which is corrected herein with respect to a number of variants of that model. Primary test domains are formed of non-pitched percussive (NPP) and pitched non-percussive (PNP) sound events. 16 detection functions are investigated, including a number of novel and recently published models. Di#erent detection functions are seen to perform well in each case, with substantially worse onset detection overall for the PNP case. It is contended that the NPP case is e#ectively solved by fast intensity change discrimination processes, but that stable pitch cues may provide a better tactic for the latter.

Citations