Arabic Isolated Word Speaker Dependent Recognition System

This paper describes the development of a new Arabic isolated word speaker dependent recognition system based on a combination of several features extraction and classifications techniques. Where, the system combines the methods outputs using a voting rule. The dataset used in this system include 40 Arabic words recorded in a calm environment with 5 different speakers. We compared 5 different methods which are pairwise Euclidean distance with Mel-Frequency cepstral coefficients (MFCC), Dynamic Time Warping (DTW) with Formants features, Gaussian Mixture Model (GMM) with MFCC, Dynamic Time Warping (DTW) with MFCC features and Itakura distance with Linear Predictive Coding features (LPC) and we got a recognition rate of 85.23%, 57% , 87%, 90%, 83% respectively. In order to improve the accuracy of the system, we tested several combinations of these 5 methods. We find that the best combination is MFCC | Euclidean + Formant | DTW + MFCC | DTW + LPC | Itakura with an accuracy of 94.39% but with large computation time of 2.9 seconds. In order to reduce the computation time of this hybrid, we compare several subcombination of it and find that the best performance in trade off computation time is by first combining MFCC | Euclidean + LPC | Itakura and only when the two methods do not match the system will add Formant | DTW + MFCC | DTW methods to the combination, where the average computation time is reduced to the half to 1.56 seconds and the system accuracy is improved to 94.56%.