Integration of Different Optimization Approaches for the Classification of Different Arabic and English Voice Commands

Abstract

Several hyperparameters represent major sensitive factors for deep learning models. For this, different hyperparameter optimization approaches are proposed to accelerate the convergence towards the optimal configurations and to support the calculations during the long learning times, and thus to give improved performance. These approaches include Bayesian Optimization (BO), Hyperband and Tree Parzen Estimator (TPE), and they are proposed for optimization task in this paper. Also, Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) are suggested as classifiers and Mel frequency cepstrum coefficients (MFCC) and Mel as features. Experiments showed that the best results in terms of evaluated performances (Precision = 94.96%, Recall = 94.85%, F1 = 94.85%) were obtained with the combination of LSTM and MFCC (LSTM-MFCC) with BO on English voice command database compared to those obtained with other combinations of features and classifiers with different optimization approaches. Moreover, there is evidence that BO converged faster than TPE and HB, and converged to better configurations.