Partial Classification in Speech Recognition Verification

Abstract

Due to speech recognition imperfections, recognition results need to be verified before being used in real-life applications. Here we present two perspectives for recognition verification: direct classification and partial classification based on confidence measures. Linear classifiers, decision trees and perceptrons are used here as direct classifiers. On the other hand, we compute confidence measures through several methods, being MLP’s and evolutionary fuzzy systems the best performing ones. Experimentation with three types of speech input reveals that higher correct verification rates can be achieved when verification is based on confidence measures. Moreover, classification rates can be improved when verification does not have to deal with “uncertain” examples, which are not classified. Partial classification represents a trade-off between verification accuracy and the number of recognition results verified.