Off-line Arabic Handwritten Recognition Using a Novel Hybrid HMM-DNN Model

In order to facilitate the entry of data into the computer and its digitalization, automatic recognition of printed texts and manuscripts is one of the considerable aid to many applications. Research on automatic document recognition started decades ago with the recognition of isolated digits and letters, and today, due to advancements in machine learning methods, efforts are being made to identify a sequence of handwritten words. Generally, based on the type of text, document recognition is divided into two main categories: printed and handwritten. Due to the limited number of fonts relative to the diversity of handwriting of different writers, it is much easier to recognize printed texts than handwritten text; thus, the technology of recognizing printed texts has matured and has been marketed in the form of a product. Handwritting recognition task is usually done in two ways: online and offline; offline handwriting recognition involves the automated translation of text in image format to letters that can be used in computer and text-processing applications. Most of the research in the field of handwriting recognition has been conducted on Latin script, and a variety of tools and resources have been gathered for this script. This article focuses on the application of the latest methods in the field of speech recognition for the recognition of Arabic handwriting. The task of handwritten text modeling and recognizing is very similar to the task of speech modeling and recognition. For this reason, it is possible to apply the approaches used for the speech recognition with a slight change for the handwriting recognition. With the expansion of HMM-DNN hybrid approaches and the use of sequential objective functions such as MMI, significant improvements have been made in the accuracy of speech recognition system.  This paper presents a pipeline for the offline Arabic handwritten text recognition using the open source KALDI toolkit, which is very well-known in the community of speech recognition, as well as the use of the latest hybrid models presented in it and data augmentation techniques. This research has been conducted on the Arabic KHATT database, which achieved 7.32% absolute reduction in word recognition error (WER) rate.