ISCA Workshop on
In this paper we present an onset detection algorithm that consists of two parts, the detection of transient peaks in an audio spectrum and the classification of the peaks, adapting a model derived from the Bayesian Theory of Surprise. The model is an unsupervised, robust adaptation of conjugate priors, providing the distributions of beliefs about the number of the transient peaks, in a time space as well as in a frequency space. The novelty points marked by the model are then classified according to their relevance in order to filter out non-onset events, caused for example by a background noise. It has been evaluated using a collection of over 170 music excerpts. Our experiments show that the new model can provide an overall performance close to the current state of the art solutions. We discuss the advantages of the presented approach and the ways to overcome its shortcomings and the possible directions of future research.
Index Terms: onset detection, Bayes, modeling, surprise, novelty
Bibliographic reference. Holonowicz, Piotr / Herrera, Perfecto (2010): "Detection of polyphonic music note onsets by application of the Bayesian theory of surprise", In SAPA-2010, 37-42.