A new method of sequential information analysis for consonant perception

Abstract

A new method of information analysis which considers all entries in the confusion matrix is proposed for identifying important articulatory features for consonant perception. For a system of N features, the important features are selected sequentially in a series of iterations. In the first iteration, the feature whose transmitted information is maximum is selected. In the second iteration, the first feature is combined with each of the remaining N−1 features to form N−1 sets of two joint features. The second feature is selected by choosing the set of joint features which has the maximum transmitted information. The transmitted information for a set is determined by summing its unconditional transmitted information and a correction factor to account for redundancy. The third and subsequent important feature are selected in the same manner. The selection process terminates when the transmitted information for the selected set of joint features in the last iteration is equal to the maximum possible transmitted information for the confusion matrix. The proposed method of analysis was tested with several confusion matrices published in past research papers, and the results showed that it was able to select a common set of important features consistently across the confusion matrices.