The Optimization And Choosing Of The Gain Filter Of The LD-CELP Algorithm

The LD-CELP algorithm with low rate and low complexity had very important meaning in the field of communication. G.728 was the only 16kbit/s ITU Recommendation in the low delay speech coding algorithms at present. The research aimed to reduce its rate and its complexity to improve G728 algorithm. In this research, the complexity of G728 could be greatly reduced, and the average segmented SNR could be improved by using the least squares adaptive algorithms to replace the Levison_Durbin algorithm. BP neural network also can improve the SNR. But the BP algorithm complexity will be hoped to decrease. This research will help others to use the network better.The voice signal is a process with the character of nonlinearand nonstationary. But the traditional speech algorithm use linear prediction algorithm to deal with the voice signal. The research improves this test result using BP neural network algorithm and the least square algorithm to replace the LevinsonDurbin algorithm. For the linear algorithm, we introduce the increasing memory recursive algorithm and the finite memory recursive algorithm into the G728 algorithm. For the nonlinear algorithm, we introduce BP neural network into the G.728 algorithm. In this research, when observing the gain filter, the quantization is not existed. Since we choose the estimation of SNR to test the better or worse of the gain filter, we can divide the gain optimization from the gain quantization, and then we get a better result with this method.In this research, we observe the effect of different algorithm for the gain filter in the G728 algorithm. Using the increasing memory recursive algorithm can decrease the algorithm complexity, and improve the effect of the gain filter. It will help to improve the gain quantization.