Dr. OMAIMA N. A. AL-ALLAF
Journal of Theoretical and Applied Information Technology
15th June 2017. Vol.95. No 11
Abstract
Speech signals are effected by noise generated by various sources of interferences. Removing noise from
speech signals can be regarded as an active research area in signal processing. Thus, we need powerful
methods in this area. Therefore, Function Fitting (FitNet) Artificial Neural Networks model was used in this paper
for enhancing speech signals. Particle Swarm Optimization (PSO) was used during FitNet learning process to optimize
the FitNet learning parameters (such as learning rate, momentum variable and network weights) to achieve best results
of speech signal enhancement. At the same time, different optimization techniques for optimizing the values of learning
parameters were suggested in this work. This is done to improve the performance of FitNet model for signal
enhancement. Better results (320 learning steps, PSNR equal 38 and mean square error (MSE) equal 0.0027) from
experiments were achieved when adopting PSO with FitNet with swarm size equal 40 and PSO number of iterations
equal 100. Good results (312 learning steps, PSNR equal 35.94 and MSE equal 0.00002) were obtained also when
adopting the suggested optimization techniques (learning rate equal 0.00003, 5 hidden units in one hidden layer with
the using of Levenberg-Marquardt (LM) as learning algorithm) for optimizing the learning parameters.