Correction of Dental Articulation Disorders in an Arabic Speaker Using Artificial Neural Networks

Abstract

The aim of our work is to correct disorders caused by wearing the denture
during the pronunciation of the 10 arabic dental phonemes
To achieve this goal, we developed a corpus of words containing these phonemes in different positions that exist in Arabic (initial, median and final). For
this, we chose two sets of subjects: Normal Subjects (NS) as Reference and a set
of Subjects Wearing Dentures (SWD).
To calculate the Recognition Rate (RR) of dental phonemes uttered by NS
and SWD, we start with data acquisition that consists of saving the selected
corpus of 10 phonemes and then by performing manual segmentation using the
Speech Filing System (SFS). The Matlab software was used to code the system components. The acoustic analysis we performed, consists of extracting the
MFCC (Mel Frequency Cepstral Coefficients) and Delta MFCC (first derivative
of MFCC). The RR of dental phonemes are performed by using Multi Layer
Neural Networks (Multi Layer Perceptrons: MLP). To solve the problems of
pronunciation caused by the wearing of dental prosthesis, we calculate the Euclidean Distance (ED) between the two categories of parameters and repeat the
phonemes to decrease up to reach the neighborhood of zero. The results are promising since a satisfactory classification rate of 75.50% was achieved when the
task consisted of verifying that problematic phonemes were effectively corrected. However, the RR can be improved by enriching the corpus and enhancing
the recording conditions.