基于动态小波神经网络非线性预测的语音编码系统

Nonlinear predictive speech coding system based on DWNN

现在,所有的语音编码系统都采用线性预测技术,但对于本质非线性的语音信号而言,线性预测是不够的.因此,本文提出一种带反馈单元的动态小波神经网络并将其应用于语音编码系统,并对其函数逼近能力和学习高维函数的优越性进行分析.由于反馈单元的内部记忆能力,动态神经网络具有对长时相关的预测能力并能在一定程度上克服小波神经网络的'维数灾难'问题;在对语音信号的预测中,动态小波神经网络预测器的预测性能很好,虽然其预测阶数很低(仅为2).由于预测器较好的预测性能,当将此预测器用于语音编码系统中的后向预测时,实验结果表明:新系统的恢复语音平均分段信噪比比ITU的G.721标准提高3～4dB但二者码率相同.另外非线性预测语音编码系统的计算量是可以接受的.

Now, linear prediction （LP） technology is popularly used in speech coding. But LP is not sufficient for non-linear speech signals. So, in this paper, we present a dynamic wavelet neural network （DWNN） with a feedback unit and try to use it for nonlinear prediction in speech coding system. Its ability to learn function and superiority in appropriating the high dimensional function are also analyzed. Because of the inner memory of feedback unit, the dynamic wavelet neural network performs well in learning long-term dependences and can overcome the ＂curse of dimension＂ problem which often occurs in wavelet networks to some degree. The nonlinear predictor based on DWNN has high prediction gain though predictive order （only 2） is low when used to speech signals. Due to good prediction, when we use DWNN in backward predictor which could not bring increase of bit rate in speech coding system, the experiment results show that： the new system has better synthesized speech and its average segmental signal-noise-rate is improved more than 3dB than ITU＇s G.721 speech coding standard system but its coding rate doesn＇t increase. Meanwhile its computational cost is acceptable.