Arabic script is inherently cursive in nature; therefore, optical character recognition (OCR) techniques developed for other scripts are generally inappropriate for Arabic. We present an OCR system for the recognition of handwritten Arabic words. Several original contributions have been made. First, we have proposed a novel feature vector to represent handwritten Arabic words based on the ellipsoid approximation of maximally stable extremal regions (MSERs). This feature vector is compact, robust, and well suited to handle writer-induced variation in handwritten script. Second, we present a new database, Saudi Arabian city name (SACN) database consisting of handwritten Saudi Arabian city names, which is the first such publicly available database. Finally, we present promising experimental results on two databases. One of the main experiments compares the proposed MSER feature vector with several well-established feature extraction methods using the proposed SACN database. The proposed feature vector performed better than all the standardized methods, with 100% correct identification rate. Another experiment is designed to compare our system to other recent studies on handwritten Arabic text recognition using the extensively employed IFN/ENIT database. Using the standard experimental protocol, our system outperformed all others with 92.64% correct identification rate.