Using MATLAB for Analysis of TRAP System

Abstract

This article describes a Matlab function for reading and processingfile outputs from a structure of classifiers. These classifiers -neural nets - are used in speech recognition based on temporaltrajectories (TRAP) of energy in frequency bands. This nonstandardapproach is introduced and the program is presented. The utility ofresulting figures is enhanced by the possibility of reading anddisplaying results from all critical bands at once. Resulting analysesare more focused on the reliability of classifiers than on the basicaccuracy measure. The first uses colors and their depth to display bothcues, reliability and accuracy, in one informative picture. Others arefocused on more precise measures, where it is possible to preciselydefine classifier mistakes.