Adaptive Gain Prediction Speech Coding System Based On Neural Network

Artificial neural network is a complicated information processing one made of many processing units. This network has the ability of learning memory and input information trait extracting. It is intelligent because it can learn from examples. Now it receives great attention and gets successful application such as mode recognize and image processing > control and optimize ^ predict -, communication etc.Speech signal is got in essence non-stationary and nonlinear. But all along, traditional speech processing method uses linear prediction. Aimed at this shortage, this paper introduces neural network in adaptive gain prediction of speech coding system, and studies the structure and learning algorithm of gain nonlinear prediction of speech coding system, which is used in G.728 algorithm.This paper studies BP network, realizes the method of gradient descent, gets better result than traditional one. In addition, this paper studies RBF network, realizes the method of orthodoxy least square, gets rapid speed than BP network by 20, and the quality of speech has a little improvement than BP. In order to realize neural network, this paper uses VC++ and MATLAB tools. The experiment results show: the sentence's average segment SNR of speech coding system based on BP network has improvement by 2dB than ITU-T G.728 standard algorithm,the gain's average segment has improvement by 3dB. In the same way, the sentence's and gain's average segment SNR of speech coding system based on RBF network has 2dB, 3.1 dB than that of ITU-T G.728 standard algorithm separately.At last, at the cost of improved SNR, the order of synthesized filter can be reduced. The result show: when the order is 10, SNR of neural network is higher than 20 order of former algorithm, and the calculation is much reduced. The analysis show: the improved algorithm lets the overall calculation of G.728 be reduced by 40%, and at the same time the quality of speech has no change.