A dynamic wavelet neural network for prediction of speech signal

A dynamic wavelet neural network (DWNN) with a feedback is described. Its approximation capacity for learning the long-term dependences between speech signals is also studied. Due to inner-memory, its predictive property is good even if the prediction order (only 3) is low. In the experiments, DWNN performs well and seems to be more promising than RNN and RBF network for speech prediction.