Speech pattern classification using Large Geometric Margin Minimum Classification Error training

Abstract

As one of the recent popular discriminative training methods, Minimum Classification Error (MCE) training aims at efficiently developing high-performance classifiers through the minimization of smooth (differentiable in classifier parameters) classification error count loss. However, MCE training, sometimes referred to as Functional Margin (FM) MCE training, does not necessarily guarantee training convergence to a high level of robustness. To solve this problem, a new version of MCE training, called Large Geometric Margin Minimum Classification Error (LGM-MCE) training, has recently been developed by introducing a geometric margin for a general form of discriminant functions for fixed-dimensional vector pattern samples. Its effectiveness in achieving robustness to unseen samples has been proven in various tasks that classify fixed-dimensional patterns. Leveraging this advance in MCE training formalization, we newly define LGM-MCE training for the classification of patterns of variable length, e.g. speech patterns, and demonstrate this training's effectiveness in a spoken-word classification task.