Automatic Arabic pronunciation scoring for computer aided language learning

• Published in: 2013 1st International Conference on Communications, Signal Processing and Their Applications pp. 1 - 6
• Main Authors: Ahmad Khan, Ali Fauzi, Mourad, O., Mannan, A. M. K. B., Dahan, H. B. A. M., Abushariah, M. A. M.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.02.2013
• Subjects: Accuracy; Arabic language; automatic pronunciation assessment; Baum Welch algorithm; Computers; Hidden Markov models; Hidden Markov Models (HMMs); Log-Likelihood probability; Mel frequency cepstral coefficient; Speech; Speech recognition; Viterbi algorithm
• ISBN: 1467328200; 9781467328203
• DOI: 10.1109/ICCSPA.2013.6487246

Automatic articulation scoring makes the computer able to give feedback on the quality of pronunciation and eventually detect some phonemes on miss-pronunciation. Computer-assisted language learning has evolved from simple interactive software that access the learner's knowledge in grammar and vocabulary to more advanced systems that accept speech input as a result of the recent development of speech recognition. Therefore many computer based self teaching systems have been developed for several languages such as English, Deutsch and Chinese, however for Arabic; the research is still in its infancy. This study is part of the "Arabic Pronunciation improvement system for Malaysian Teachers of the Arabic language" project which aims at developing computer based systems for standard Arabic language learning for Malaysian teachers of the Arabic language. The system aims to help teachers to learn the Arabic language quickly by focusing on the listening and speaking comprehension (receptive skills) to improve their pronunciation. In this paper we addressed the problem of giving marks for Arabic pronunciation by using a Automatic Speech Recognizer (ASR) based on Hidden Markov Models (HMM). Therefore, our methodology for pronunciation assessment is based on the HMM log-likelihood probability, however our main contribution was to train the system using both native and non native speakers. This resulted on improving the system's accuracy from 87.61% to 89.69%.