The effects of distortion measures and feature sets on neural network classifiers

The authors investigate the use of two types of neural networks, multilayer perceptrons (MLP) and learning vector quantizers (LVQ), as applied to isolated speaker-independent vowel recognition as a typical classification task. The LVQ algorithm used is a modification called the frequency-sensitive competitive-learning (FSCL) LVQ. The performance of each of these networks for different input feature sets is evaluated and compared. The effects of different distortion measures on recognition performance are also studied. The results show that the choice of the input feature set and the distortion measure can significantly affect recognition performance. It is shown that, while both the backpropagation (BP) and FSCL-LVQ algorithms can be applied to a set of vowel-recognition tasks, the FSCL-LVQ procedure offers an advantage over the MLP approach. The FSCL-LVQ algorithm allows the use of any appropriate distortion measure for particular input features, while the BP algorithm optimizes the weights by minimizing the squared errors between the actual and desired outputs. Consequently, for some tasks, the LVQ architecture can perform more accurate classification