Automatic Recognition, Investigation, and Analysis of the Spoken Arabic Alphabet

Abstract

Alphabet recognition is needed in many applications for retrieving information associated with the spelling of a name, such as in airline reservations and addresses. This is a difficult recognition task due to the acoustic similarities existing between alphabets in the same set of alphabet (e.g., the English E-set and the Arabic Baa-set). In this paper the Arabic alphabets were investigated from the Automatic Speech Recognition (ASR) problem point of view. The Hidden Markov Model (HMM) technique was used in this research due to its high performance compared to other strategies. As a way of implementing this technique, the hidden Morkov model toolkit (HTK) was used to design an isolated word recognizer with phoneme based HMM models. In addition to the analysis of Arabic alphabets with reference to time and frequency domains, two experiments were conducted. A database with a size of 14,500 tokens was created through reliance on fifty Arabic native speakers. The spoken Arabic alphabet has more than one set, each of which contains alphabets that are acoustically very similar. The biggest set is the Baa-set, which, in our experiment, caused many errors in our system .The designed alphabet recognition system achieved83.34% correct for alphabet. The spoken alphabet “Hamzah” got almost 100% recognition rate, but,relatively, bad performances were encountered with alphabets like “Baa”, “Taa”, “Thaa”, “H_aa”,“Thaal”, “Raa”, “Seen”, “T_aa”, “Dhaa”, and “Faa”. The worst accuracy was encountered with theArabic spoken alphabet “Faa”, which had an accuracy of 31.8%.