Confidence Measures in Multiclass Speech Emotion Recognition Using Ensemble Learning To Catch Blunders

Author(s):

Alan Murphy , National University Of Ireland Galway; Sam Redfern, National University Of Ireland Galway

Keywords:

Confidence Measures, Emotion Recognition, Ensemble Learning, Multiclass Speech Classification

Abstract:

Although Speech Emotion Recognition (SER) systems have continually been improved with regard to outputting decisions on class membership, it is only very recently that Confidence Measures (CM) have been incorporated into such systems. It is easy to presume that classifiers such as k-Nearest Neighbor, Naïve Bayes or Support Vector Machines that can readily output numeric distributions, can therefore easily produce confidence estimates. However these numeric outputs have proven not to be well correlated with classification confidence. With this in mind the contribution of this paper is threefold, 1) providing the first successful demonstration of confidence extraction in the multiclass problem of SER by using metrics from a k-NN meta-classifier within an ensemble architecture . 2) We also demonstrate how this system has shown improvement on a similar approaches discussed later, and 3) finally we note how confidence measures can be useful as a tool for catching blunderous predictions made in SER problems, which can be very useful in real world deployed systems.

Other Details:

Manuscript Id	:	IJSTEV2I3013
Published in	:	Volume : 2, Issue : 3
Publication Date	:	01/10/2015
Page(s)	:	118-122

Download Article