Automatic Diagnosis of Distortion Type of Arabic /r/ Phoneme Using Feed Forward Neural Network

Abstract

The paper is not for recognizing normal formed speech but for distorted speech via examining the ability of feed forward Artificial Neural Networks (ANN) to recognize speech flaws. In this paper we take the Arabic /r/ phoneme distortion that is somewhat common among native speakers as a case study.To do this, r-Distype program is developed as a script written using Praat speech processing software tool. r-Distype program automatically develops a feed forward ANN that tests the spoken word (which includes /r/ phoneme) to detect any possible type of distortion. Multiple feed forward ANNs of different architectures have been developed and their achievements reported. Training data and testing data of the developed ANNs are sets of spoken Arabic words that contain /r/ phoneme in different positions so they cover all distortion types of Arabic /r/ phoneme. These sets of words were produced by different genders and different ages.The results obtained from developed ANNs were used to draw a conclusion about automating the detection of pronunciation problems in general.Such computerised system would be a good tool for diagnosing speech flaws and gives a great help in speech therapy. Also, the idea itself may open a new research subarea of speech recognition that is automatic speech therapy.

Keywords: Distortion, Arabic /r/ phoneme, articulation disorders, Artificial Neural Network, Praat