The extended Kalman filter algorithm for improving neural network performance in voice recognition classification

摘要

In the previous study, we have been investigated that Mel-Frequency Cepstral Coefficient (MFCCs) is very powerful tools as a feature extraction method in the voice recognition application. The MFCCs is also very suitable to work with the neural network (NN) as voice classification algorithm due to the number of MFCCs representing the voice data. In this paper, we focused in improving performance of neural network in voice recognition classification using the extended Kalman filter (EKF) as the training algorithm. Using the EKF for the training NN is proved gives excellent convergence performance in many applications. Simulation of the Backpropagation algorithm is also presented. The simulation result shows that in the training data the EKF provides the performance rates 100% and requires only 1 up to 4 epochs, while the Backpropagation provides the performance rates until 93.83% and requires 20 up to 100 epochs. In the testing data, the EKF provides the performance rates until 92% and the Backpropagation provides the performance rates until 90%. These results show that the EKF could improve the NN performances.