Adaptive networks, dynamical systems, and the predictive analysis of time series speech analysis

The authors attempt to illustrate that adaptive network techniques provide an efficient mechanism for extracting qualitative details concerning the statistics and dynamics of transient time series based on a limited amount of information. Although the reported results on speech waveforms were obtained using a traditional multilayer perceptron structure (with linear output units), very similar results were obtained with radial basis function networks with spherical Gaussian nonlinearities at the hidden units. They suggest that the observed structure is characteristic of the data itself, as opposed to an artifact of the particular network used to model the observation sequence. They also suggest that this approach indicates fruitful possibilities for coding applications