Complementary work to that carried out by Chu and Bose (1998) into speech s
ignal prediction using feedforward neural networks is reported. Results of
prediction performance and network efficiency using a recurrent neural netw
ork (RNN) are presented and compared to those obtained using a multilayer p
erceptron (MLP) and a linear predictor. An RNN structure is described which
has much better performance than that of an MLP at almost the same network
efficiency.