Speech Nonlinear Predictor with Multi-Layer Perceptron
Zhou Zhijie, Pla Uni · Journal of PLA University of Science and Technology · 2001
A speech nonlinear predictor with Multi-Layer Perceptron (MLP) is demonstrated in this paper. Compared with existing nonlinear speech predictors, two innovations are embodied in the predictor. One is the modification of mean square error function where weight regularization is used to ease overfitting. The other is the Linearized Optimization Layer by Layer (LOLL) algorithm which speeds up MLP's training procedure. Experimental results show us that: the prediction SNR of the proposed nonlinear predictor is about 2 dB higher than that of linear one at the same prediction order. A speech nonlinear predictor can practice both short term prediction and long-term prediction at the same time. Weight regularization adds 0.35 dB to prediction SNR.