Recurrent neural network speech predictor based on dynamical systems approach

A nonlinear predictive model of speech, based on the method of time delay r econstruction, is presented and approximated using a fully connected recurr ent neural network (RNN) followed by a linear combiner. This novel combinat ion of the well established approaches for speech analysis and synthesis is compared with traditional techniques within a unified framework to illustr ate the advantages of using an RNN. Extensive simulations are carried out t o justify the expectations. Specifically, the network's robustness to the s election of reconstruction parameters, the embedding time delay and dimensi on, is intuitively discussed and experimentally verified, In all cases, the proposed network was found to be a good solution for both prediction and s ynthesis.