An Efficient Mispronunciation Detection System Using Discriminative Acoustic Phonetic Features for Arabic Consonants

摘要

Mispronunciation detection is an important component of Computer-Assisted Language Learning (CALL) systems. It helps students to learn new languages and focus on their individual pronunciation problems. In this paper, a novel discriminative Acoustic Phonetic Feature (APF) based technique is proposed to detect mispronunciations using artificial neural network classifier. By using domain knowledge, Arabic consonants are categorized into two groups based on their acoustic similarities. The first group consists of consonants having similar ending sounds and the second group consists of consonants with completely different sounds. In our proposed technique, the discriminative acoustic features are required for classifier training. To extract these features, discriminative parts of the Arabic consonants are identified. As a test case, a dataset is collected from nativeon-native, male/female and children of different ages. This dataset comprises of 5600 isolated Arabic consonants. The average accuracy of the system, when tested with simple acoustic features are found to be 73.57%. While the use of discriminative acoustic features has improved the average accuracy to 82.27%. Some consonant pairs that are acoustically very similar, produced poor results and termed as Bad Phonemes. A subjective analysis has also been carried out to verify the effectiveness of the proposed system.