Investigation of some parameters of a neuro-fuzzy approach for dynamic sound fields visualization

Abstract

The present paper presents detailed investigation of some parameters of our recently proposed approach for multidimensional data clustering aimed at dynamic sound fields' visualization. These include the following: number of direction selective cells (MT neurons) applied as filters at the first step of feature extraction from the raw data; size of ESN reservoir used at the second step for feature extraction; selection criteria for proper 2D projection of the original multidimensional data; number of clusters into which data are separated. The tests were performed using real experimental data collected by a microphone array (called further “acoustic camera”) build from 18 microphones placed irregularly on a wheel antenna with a photo camera at its center. Using our approach we created dynamic “sound pictures” of the data collected by acoustic camera and compared them with the static “sound picture” created by the original software of the equipment. During investigations we also discovered that our algorithm is able to distinguish among two sound sources - a task that was not that well performed by the original software of the acoustic camera.