Duration Models for Arabic Text Recognition Using Hidden Markov Models

We present in this paper a system for recognition of printed Arabic text based on hidden Markov models (HMM). While HMMs have been successfully used in the past for such a task, we report here on significant improvements of the recognition performance with the introduction of minimum and maximum duration models. The improvements allow us to build a system working in open vocabulary mode, i.e., without any limitations on the size of the vocabulary. The evaluation of our system is performed using HTK (hidden Markov model toolkit) on a database of word images that are synthetically generated.