Alternative phonetic class definition in linear discriminant analysis of speech

The class definition in Linear Discriminant Analysis (LDA) applied in Automatic Speech Recognition (ASR) is a still discussed problem, because some issues have not been completely answered. It is not always obvious how to apply LDA in the specific ASR front-end and how to appropriately define the classes to be discriminated. In this paper we present an alternative class definition method for LDA based on phonetic segmentation. Several selected properties of phonemes such as slope and duration are used to define the alternative classes. These ´phoneme states´ are modeled separately according to their duration and occurence in the database. Comparisons against the LDA based on conventional class definition are given. Different lengths of supervectors are used to investigate the influence of the contextual information to the final performance. Several experiments with various configurations using two databases on phoneme-based continuous speech recognition task were performed. Experimental results show, depending on the technique, that the proposed method achieves comparative results compared to the conventional method.