新型的神经网络线性预测语音编码算法

提出一种新型的神经网络线性预测编码算法.针对目前自相关法存在着预测系数解误差以及协方差法存在解不稳定的缺点,算法利用最小均方准则思想显著提高了短时平均误差精度.通过窄带信道将低速率语音编码远距离传输是多媒体语音技术中的重要研究内容,采用语音信号压缩处理是解决低速率传输的有效方法之一,而线性预测编码(LPC)技术是语音压缩参数编码技术的重要内容.从线性预测编码技术入手分析和研究LPC编码技术的原理,阐述了利用最小均方准则思想来提高短时平均误差精度的方法,并介绍了预测系数的自相关求法.最后通过语音合成实验验证了该新型算法既提高了系数解的精度,又保证了系统的稳定性.在该算法下预测系数的均方误差比传统的自相关法误差减小20%以上,而且当原始信号频率较高时语音合成的效果更明显,将更加精确地复现原始语音信号.

A new algorithm for linear prediction coding based on neural network

A new algorithm of linear prediction coding based on neural network is presented.In contrast with the prediction coefficients error of self-correlative method and instability of covariance method,this algorithm improves the short-time average precision observably.As the main content of multimedia information compressing technology,long-distance transmission of the speech parameter coding by narrow band signal channel plays an important role.Speech signal compressing is an important section of multimedia information compressing technology.The theory of LPC coding is researched based on the linear prediction coding technology,the least mean square rule is applied for improving the short-time average precision,and the self-correlative calculation of prediction coefficients are introduced.Finally the speech synthesizing experiment is carried out and the result demonstrates that it is not only proved to be an effective way that improves the precision but also ensures systemic stability.Up to 20 percent decrease of the mean square error of prediction coefficient by our method is obtained in contrasting with that of a traditional self-correlation method.When the original signal has a higher frequency,the quality of speech synthesizes with applying this algorithm would be more effective.