Improving LD-aCELP's Gain Filter

摘要

LD-aCELP algorithm has 2.5ms delay and speech coding rate is 8Kbit/so The adaptive codebook and backward pitch detection is used. LD-aCELP depends on the Levinson-Durbin (L-D) algorithm to update gain filter coefficients. Because quantizer has not existed at optimizing gain filter, the quantization SNR can not be used to evaluate its performance. We use a new scheme to estimate SNR so that the gain predictor can be separately optimized with the quantizer. In this paper, using this scheme L-D method is replaced by three different methods which are the weighted L-S recursive filter, the finite memory recursive filter and the BP neural network, respectively. Experiments showing, they are all very effective to improve gain filter performance. The weighted L-S algorithm has the best effect, which is accordant with real speech coding. Its average segment SNR is higher than LD-aCELP about 0.720dB.