A segmentation algorithm based on acoustical features using a self organizing neural network.

Abstract

In this paper we investigate the use of a self-organizing map in an acoustic segmentation task. The aim is to obtain a limited number of acoustic classes and to segment whenever a change in the class between two adjacent frames occurs. Energy in different frequency ranges is used as input in the map training process. A structure based on a Kohonen map connected to a neural network trained with the back-propagation algorithm is proposed.