Feature selection using game theory for phoneme based speech recognition

Reduced feature set containing relevant features for identifying individual phonemes were obtained using two game-theoretic formulations. In one formulation feature selection algorithm tries to obtain features that maximize the accuracy of the classifier, and in another it obtains features that minimize the misclassification rate of the classifier. Experiments are run on the TIMIT database for generating classifiers using the reduced feature set obtained from our feature selection algorithms and compared against classifiers generated using all of the features. The results show that, classifiers generated using the reduced feature set out performed classifiers generated from all of the features. In addition, reduced feature sets obtained using proposed feature selection algorithms could significantly reduce storage and computational complexity without compromising on accuracy of classifiers.