Classification of Spanish vowels and digits using neural networks

In this paper we describe the use of Multilayer Perceptron Array for learning and classifying speech signals, using characteristic vectors of reconstructed dynamics. First, we consider the phonatory system as a black box, where the only available data is its output: the speech signal. This is a way of accessing underlying dynamics, and is the starting point for two kinds of experiments: classification of vowels and digits, with Venezuelan Spanish voices. Results verify positively that characteristics vectors extracted from underlying dynamics hold discriminative power for distinguishing between classes of speech signals. Besides, neural networks are able to generalize using this kind of data.