An intuitive class discriminability measure for feature selection in a speech recognition system

In this paper we propose an intuitive method for feature selection and dimensionality reduction of a set of selected features. We analyze the feature set by applying an intuitive class discriminability measure. For this purpose, we use the proposed schemes for the “editing” of the training datasets. Editing algorithms basically smooth the decision boundaries and improve the performance of the classification with finite sample size. We propose that an intuitive measure for class separability or the overlap between classes can be the “number of remaining points” after editing the dataset using an algorithm with an idea similar to Wilson-editing algorithm over the total number of points. The algorithm that we aim to apply is a modification of the Wilson-Gabriel editing algorithm. We apply this method to compute the overlap between phonetic classes for two different feature sets, Mel Frequency Cepstral Coefficients (MFCCs), and Linear Prediction Coefficients (LPCs). These feature sets are commonly used in speech recognition systems. The proposed measure indicates a higher class discriminability for MFCC feature space comparing to LPC feature space.