Nonlinear prediction of speech signals using memory neuron networks

The authors present a feed-forward neural network architecture that can be used for nonlinear autoregressive prediction of multivariate time-series. It uses specialized neurons (called memory neurons) to store past activations of the network in an efficient fashion. The network learns to be a nonlinear predictor of the appropriate order to model temporal waveforms of speech signals. Arrays of such networks can be used to build real-time classifiers of speech sounds. Experiments where memory-neuron networks are trained to predict speech waveforms and sequences of spectral frames are described. Performance of the network for prediction of time-series with minimal a priori assumptions of its statistical properties is shown to be better than linear autoregressive models