A recurrent neural network speech predictor based on dynamical systems approach

A nonlinear predictive model of speech, based on the method of time delay reconstruction, is presented and approximated using a fully connected recurrent neural network (RNN) followed by a linear combiner. This novel combination of the well established approaches for speech analysis and synthesis is compared to traditional techniques within a unified framework to illustrate the advantages of using an RNN. Extensive simulations are carried out to justify the expectations. Specifically, the networks' robustness to the selection of reconstruction parameters, the embedding time delay and dimension, is intuitively discussed and experimentally verified. In all cases, the proposed network was found to be a good solution for both prediction and synthesis.