Nonlinear mapping function of the HCNN(Hidden Control Neural Network) can change over time to handle the temporal variability of a signal like speech by combining nonlinear prediction of conventional neural networks with the segmentation capability of HMM. Weighted distance is a statistical distance measure with weights equal to the inverse-variance of the components of feature vectors. In this thesis, HCNN with its error measure given by weighted distance is proposed and the performance of this system is evaluated.
The recognition accuracy of 97.35% is obtained for speaker-independent Korean digit recognition experiment. The performance improvements of the proposed system over HCNN with Euclidean distance and the HMM are about 2.3% and 3.5% respectively.
