Abstract:
Deafness is one of disabilities that spread around the world, over 5% of the world’s population has disabling hearing loss. The majority of people with disabling hearing loss live in low and middle income countries. It has many impacts functional, social, emotional, and economic impact.
According to the meeting of Arab deaf real state and difficulties the statistic of deaf people in Arab world about 15 million, most of them not educated or with primary education; due to less educational politics in the area. Thus the research is attractive, computerized learning method which help fighting illiteracy among deaf people generally and Arabian deaf specially and to keep them in continuous learning process.
The aim of designing Arabic phonetics recognition system using artificial neural networks is to recognize Arabic phonemes to assist deaf people, others who suffer from some disabilities even healthy people to gain their time and effort, by converting Arabic phonetics into Arabic text. Arabic phoneme recognition is done through signal acquisition, preprocessing, feature extraction, classification and recognition using pattern recognition neural networks technique.
To achieve these objectives Matlab was used which has powerful capability to identify analysis and recognize speech signal in fast manner through signal processing toolbox, and neural network toolbox.
The algorithms which were implemented include firstly signal preprocessing in which the speech signal transformed into digital form. Three steps of signal preprocessing have been applied, normalization, filtering, and windowing. Secondly feature extraction in which three types of speech parameters were extracted (energy, fast Fourier transform, and formant estimation with linear predictive coefficients). Lastly supervised pattern recognition networks for classification. Three networks were built. The network trained with scaled conjugate gradient backpropagation.
The implemented Matlab algorithms achieved the expected results. Pattern recognition neural networks for classification and recognition achieved average recognition rate of 96.17% to all 96 recorded phonemes. The outputs of simulated network in both binary and written Arabic letter with short vowel, correspondent to each Arabic phoneme.
