An application of discriminative feature extraction lo filter-bank-based speech recognition

A pattern recognizer is usually a modular system which consists of a featur e extractor module and a classifier module. Traditionally, these two module s have been designed separately, which may not result in an optimal recogni tion accuracy. To alleviate this fundamental problem, the authors have deve loped a design method, named Discriminative Feature Extraction (DFE), that enables one to design the overall recognizer, i.e., both the feature extrac tor and the classifier, in a manner consistent with the objective of minimi zing recognition errors. This paper investigates the application of this me thod to designing a speech recognizer that consists of a filter-bank featur e extractor and a multi-prototype distance classifier. Carefully investigat ed experiments demonstrate that DFE achieves the design of a better recogni zer and provides an innovative recognition-oriented analysis of the filter- bank, as an alternative to conventional analysis based on psychoacoustic ex pertise or heuristics.