The main objective of this paper is to recognize isolated Arabic words using segmentation. In order to achieve this objective, a hierarchical classification system has been designed to classify Arabic phonemes. The first stage of this system is to discriminate between vowels, voiced consonants and unvoiced consonants with a classification rate of 94 %. Then, the voiced consonants are classified into voiced fricatives, voiced plosives or vowel-like phonemes with a classification rate of 85 %. The set of unvoiced consonants is also divided into unvoiced fricatives and unvoiced plosives with a rate of 87 %. The labeled frames are further classified into different phonemes. Contextual smoothing is used to correct some misclassified frames. Thus, instead of storing the features of each word, the features of each phoneme are stored using much smaller storage area.