语音编码算法的神经网络增益滤波器  

The Neural Network for Gain Filter in Speech Code Algorithm

作　　者：张刚[1] 谢克明[1] 赵哲峰[1] 薛春雨[1] 

机构地区：[1]太原理工大学信息工程学院,山西太原030024

出　　处：《测试技术学报》2006年第3期246-250,共5页Journal of Test and Measurement Technology

基　　金：国家自然科学基金资助项目(60372058);山西省自然科学基金资助项目(20041046)

摘　　要：增益-波形乘积码书结构广泛用于CELP语音编码算法,它们使用L ev inson-Durb in(L-D)方法更新增益滤波器系数.本文对BP神经网络算法与L-D方法进行了比较.用BP神经网络增益滤波器进行语音编码,其计算量仅为G.728的L-D方法的6.7%,但平均分段SNR高出G.728算法0.156 dB.同时,用BP神经网络算法评价了16和20样点激励矢量增益滤波器,效果同样很好.但是,由于考察增益预测器时量化器还不存在,因此无法用量化信噪比评价滤波器性能.本文提出一种信噪比估计方法,可使增益预测器的优化与量化问题分开处理.实验表明用这种信噪比估计方法选择增益滤波器十分有效.The structure of the gain-shape product codebook has been used to almost all CELP speech coding algorithm. Their gain filter coefficients were updated depending on the Levinson-Durbin （L-D） method which is contrasted with the BP neural network gain filter proposal designed in this paper. Using the BP neural network filter, the calculation quantity is only 6.7 percent of the L-D method＇s and its average segment SNR is about 0. 156dB higher than the G. 728. It is also used to evaluate the case that excitation vector is 16 and 20 samples respectively, the BP neural network algorithm has similarly good result. Because quantizer has not existed at optimizing gain filter, the quantization SNR can not be used to evaluate its performance. This paper proposed a novel scheme to estimate SNR so that the gain predictor can be separately optimized with the quantizer. Experiments show that it is very effective using this method of evaluating gain filter performance.

关 键 词：G.728 增益滤波器 信噪比估计 BP神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]