|International Journal of Computer Applications
|Foundation of Computer Science (FCS), NY, USA
|Volume 184 - Number 37
|Year of Publication: 2022
|Authors: Azza E.B. Ibrahim, Rana S.M. Saad
Azza E.B. Ibrahim, Rana S.M. Saad . Intelligent Categorization of Arabic Commands Utilizing Machine Learning Techniques with Short Effective Features Vector. International Journal of Computer Applications. 184, 37 ( Nov 2022), 25-32. DOI=10.5120/ijca2022922440
Different technologies are now being employed to improve the quality of life, particularly for the disabled and elderly. Speech is the quickest and most convenient method of communicating with people and technology. The majority of the works have focused on English speech; however, there is some interest in Arabic. In this study, an Arabic dataset is created, which will eventually be used to control a mobile assistant robot. Arabic is a challenging language to learn because of its many dialects, each of which has its own impact on the spoken word. The Egyptian Arabic Speech Commands (EASC) dataset was compiled from people of different backgrounds, ages, and genders who spoke in colloquial dialects. The Arabic recognition test is made more difficult by this fluctuation. Using various machine learning techniques, Arabic speech commands were classified. Mel Frequency Cepstral Coefficients were used to create an effective feature vector (MFCC). Spectral centroids and signal power are combined with MFCC to generate an enlarged features vector, which improves recognition accuracy. Because these commands will drive a robot in real time, they must be classified quickly. As a result, the training features vector's dimension is lowered by performing some statistical calculations on it. Support Vector Machines (SVM), Random Forest decision tree (RFT), Neural Network (Multi-Layer Perceptron, MLP), and k-nearest neighbors (KNN) approaches were employed as intelligent classifiers. A thorough examination of the classifiers' various parameters was carried out. With a classification accuracy of 94.84 percent, the SVM approach outperformed other techniques. We concluded in this research that the enlarged features vectors with a lower dimension are more effective for this problem and can be employed in real applications.