Abstarct: Age estimation baxsed on human's speech features, is a considerable subject in automatic speech recognition (ASR) systems. Although some researches have been done in speaker age estimation, but more studies, especially in Farsi Language speech is required. Like other speech processing systems, we are face with two main challenges in age estimation: finding an appropriate method for feature extraction, and choosing a reliable classification method. The main goal of this research is using Root Mel Ferequency Cepstral Coefficients features in an age estimation system and finding an optimum root (gama)to have a less error; and also comparing the performance of an age estimation system when using this feature, with other common features like MFCC, PLP and LPC. For feature extraction, the whole of speech signal is separated to phonemes (smallest part of speech). these phoneme signals is used for feature extraction and classification. Linear classification and Mahalanobis distance method are used for classifying each phoneme signal in one of the age groups. Expriments which has been done on FARSDAT databaxse, shows that at gama=0.006 (the root in RootMFCC method) there is the least error rate (28.69 %). besides, separating speech signal to its phonemes and extracting features from each phoneme signal, improves the results and decreases the error rate.