Isolated Persian digit recognition using a hybrid HMM-SVM

This paper introduces a new method for solving a traditional problem in isolated digits recognition in Persian language. The problem arises from pronunciation similarity of some Persian digits that are composed of very similar phonetic and spectral components. The process of recognition introduced here consists of three stages. First, the word is decomposed into small parts using efficient algorithms in order to make its Hidden Markov Model (HMM). Subsequently, based on this model, the most relevant candidates are chosen and introduced to a support vector machine (SVM) based recognizer. At the final stage, the recognition is finalized by the SVM, with the aid of a novel idea that is to segment the input word and find an entry with the maximum number of similar segments. Experimental results show that the proposed method significantly improves both the recognition accuracy and the computational complexity in isolated Persian word recognition systems.