Hierarchical phoneme recognition using node-wise relevance-optimizedfFeatures

Abstract EN

In this paper, a hierarchical phoneme recognition system is proposed.

The hierarchical approach is applied here to recursively partition the recognition problem into smaller and smaller sub-problems those are independently handled at the distinct nodes of the hierarchy.

The nodes are individually set to characterize different properties of the input phoneme, or more precisely to make separate decisions on its pertinence to the different reference subgroups of phonemes.

The full characterization of the input phoneme is achieved by traversing some root-to-leaf path through the hierarchy.

The relationships between the different features of phonemes and their pertinence to the different reference subgroups are to be objectively characterized and optimized here.

This involves specifying the decisive subset of features for each pertinence decision and neglecting the remaining features those are irrelevant to (or probably have negative effect on) that decision, at each node of the hierarchy.

The optimization applied through the feature election process here, is not aimed at reducing the amount of features to be used in the recognition process, for the purpose of decreasing the time-complexity of the system, but, is interested in enhancing the decision making accuracy of the system by avoiding the misleading features.