Automatic Identification of articulation disorders for Arabic children speakers

Abstract

Automatic identification of articulation disorders in children’s speech is very important for the diagnosis and monitoring of speech therapy. In this work, acoustic features (MFCC) have been used with the two most commonly used classification techniques in the speaker and language identification area, GMM-UBM and I-vector, for identifying three types of articulation disorders associated with phoneme [r] from Arabic children’s speech. The sound [r] has been selected as it is the most common pronunciation problem that children suffer from. The impact of [r] location in a word on the speech disorders has been investigated by considering words with [r] in the beginning, middle and end We achieved up to 75% accuracy with our I-vector system and 61% for our GMM-UBM system. Performance of these two systems are improved to 92.5% and 83.4%, respectively, when disorder classes are combined into one class