In this paper, a hierarchical Arabic phoneme recognition system is proposed in which Mel Frequency Cepstrum Coefficients (MFCC) features is used to train the hierarchical neural networks architecture. Here, separate neural networks (subnetworks) are to be recursively trained to recognize subsets of phonemes. The overall recognition process is a combination of the outputs of these subnetworks. Experiments that explore the performance of the proposed hierarchical system in comparison to non-hierarchical (flat) baseline systems are also presented in this paper.