Large-vocabulary speech recognition systems require the definition of recognition units. These units should be trainable, well-defined, and insensitive to context. In this paper a set of basic phonetic units, for the Arabic language, has been designed and each individual phone is represented by a Hidden Markov Model. The proposed set of models were tested on a speaker-trained, speech recognition task with a vocabulary of 500 basic Arabic words. The analysis of the errors made by the recognizer resulted in the addition of some context-dependent to the initial phonetic set.