Selection of optimal features for phoneme recognition in different phonetic environments

The authors introduce a modified approach to phoneme recognition. The new method is based on the assumption that efficient features for phoneme identification in connected speech, depend on the examined phoneme and on the surrounding phonemes. Hence in a given phonetic environment, each phoneme si best separated from other phonemes by means of an `individual´ feature set. This leads to a quadratic discriminant. Using the new approach an improvement of 8% in consonant recognition was demonstrated. A labeling algorithm for connected speech recognition is proposed. It employs accumulated information concerning the phonetic environment prior to consonant labeling