Abstract A new architecture for statistical classification of hand-printed characters is presented. It is based on standard preprocessing and three feature types, containing geometrical information on the position of the pixels, on the contour orientation and on the bending points, respectively. Two feature vectors at a time are used as inputs to a multi-layer perceptron-based classifier, giving rise to three simple classifiers operating in parallel. The outputs of the three different classifiers are mixed by a final supervisor realized by a perceptron layer. The overall network has been trained using digits, upper and lower case letters of the NIST Special Database 3. Classification results of the NIST Test Data 1 are provided. The system has an error rate of 2.59% on the digits of NIST Test Data 1 at zero rejection rate, while it has 2.99 and 11.00% error rate on the upper and lower case letters, respectively.