Minimum bayes error feature selection

Abstract

We consider the problem of designing a linear transformation θ ∈ ℝp×n, of rank p ≤ n, which projects the features of a classifier x ∈ ℝnonto y = θ ∈ ℝpsuch as to achieve minimum Bayes error (or probability of misclassification). Two avenues will be explored: the first is to maximize the θ-average divergence between the class densities and the second is to minimize the union Bhattacharyya bound in the range of θ. While both approaches yield similar performance in practice, they outperform standard LDA features and show a 10% relative improvement in the word error rate over state-of-the-art cepstral features on a large vocabulary telephony speech recognition task.