Unsupervised classification of audio signals by self-organizing maps and Bayesian labeling

Abstract

Audio signal classification consists of extracting some descriptive features from a sound and use them as input in a classifier. Then, the classifier will assign a different label to any different sound class. The classification of the features can be performed in a supervised or unsupervised way. However, unsupervised classification usually supposes a challenge against supervised classification as it has to be performed without any a priori knowledge. In this paper, unsupervised classification of audio signals is accomplished by using a Probabilistic Self-Organizing Map (PSOM) with probabilistic labeling. The hybrid unsupervised classifier presented in this work can achieve higher detection rates than the reached by the unsupervised traditional SOM. Moreover, real audio recordings from clarinet music are used to show the performance of our proposal. © 2012 Springer-Verlag.