It is very important to enhance the recognition accuracy of the Arabic spoken letters. The accuracy of recognition system is affected by the feature extraction and the used classifier. An effective and robust method is proposed to evaluate speech feature to improve the performance the recognition accuracy. This work introduces applying the mel frequency cepstral coefficient (MFCC) to extract the speech features. Hidden Markov model and neural network are used as classifier tools. The objective of the proposed system is to enhance the performance by introducing three systems which are proposed to recognize the spoken Arabic letters. The first is based on neural networks. The second is based on hidden Markov model. While third system is based on combination between neural networks and hidden Markov models. The accuracy of neural network is found to be 42% with MFCC for 84 spoken letters. The hidden Markov models are statistical based approach. Its performance is found to be 98.5%. But for combination system based on neural network and hidden Markov models, the accuracy of 99.25% is obtained.