The Hopfield and Hamming Networks Applied to the Automatic Speech Recognition of the Five Spanish Vowels

Abstract

In the present work, two trained classifiers in Neural Networks, specifically the Hopfield network and the Hamming network, were applied to a problem in speech recognition and the results were compared. The problem is that of the automatic recognition of the five Spanish vowels in a multi-speaker environment. There is thus a twofold purpose to the present note: to give an application of two NN algorithms and to give two different, although similar, methods for recognizing those vowels.

The results showed that Neural Networks have a good prediction capability. The Hamming network has a series of evident advantages over the Hopfield network, although the results were quite similar. The main advantage of the Hamming network, in our case, is that it has far fewer connections than the Hopfield one; while Hamming needs Np connections, Hopfield needs N2.

The models of unsupervised neural networks with the use of back-propagation algorithms will be developed in forthcoming papers. Although the success rate of the Hopfield and Hamming networks is somewhat less than that of the multi-layer perceptron network, the present models of neural networks are readily installable and easy to use, since they do not require the high level of training necessary for the multilayer perceptron.

The algorithms described in this paper have been written on Macintosh computers, with a microphone/digitizer arrangement which provides sampling up to 22kHz. It was developed with a mathematical signal laboratory using parameters in the realm of the frequency spectrum.