Phonetic recognition of arabic latters using neural networks

This work is a trial to recognize the Arabic letters spoken by any speakera step to convert the spoken Arabic word to text. A suggested speaker-dependent Arabic letters recognition system is discussed. The system isased on the short time energy per frame as the feature extraction and theynamic Time Warping (DTW) program to align the analogous framesom different samples of the same signal. Then giving this feature to theeural network for training on it and testing the accuracy of recognition.e best recognition accuracy obtained for 1680 samples Arabic letterscollected from 12 speakers (6 male speakers and 6 female speakers) is84.31 % for 560 test samples and 96.43% for the learning samples. Also thefeatures based on cepestrum and LPC coefficient are tested, but therecognition accuracy was poor (about 46.33% and 31.4% for usingcepestrum and LPC feature respectively). So a classification of data intogroups is carried out according to the similarities between different lettersusing the three features to improve the recognition accuracy. The averagerecognition accuracy is improved to 65.38% when using LPC feature, therecognition accuracy is improved to 79.18% when using cepestrum feature,and the recognition accuracy by using the energy per frame feature isimproved to a good percentage 98.92%.A suggested neural networks system for improving the recognitionaccuracy of the isolated Arabic letters spoken by any speaker (male orfemale) is introduced. This system is based on the energy feature and theneural network used for classification or recognition from the type ”AllClasses One Network (ACON)”. In an ”ACON” network a single neuralnetwork has responsibility for recognizing all characters and only onesingle network needs to run to achieve recognition. By using the proposedsystem, the recognition accuracy is increased from 84.31 % to 98.92%.This is achieved by classifying all the input data of all letters into fivedifferent groups and each group containing a definite number of letters.