Creating a lexicon for a speech recognizer that uses word models based on phonemes can be a difficult task when the lexicon of the recognizer must contain proper nouns or uncommon words. Recently, researchers have developed techniques for automatic pronunciation learning in which the word models are based on data-driven subphonetic units. These units are able to model acoustic events at a much finer level of detail. In this paper we explore the use of these new techniques within our hybrid MLP/HMM speech recognizer.