【作者】 赵峰；

【导师】 张刚；

【作者基本信息】 太原理工大学 ， 信号与信息处理， 2005， 硕士

【摘要】 人工神经网络是采用大量的处理单元连接起来构成的一种复杂的信息处理网络。这种网络具有与人脑相类似的学习记忆能力和输入信息特征抽取能力。人工神经网络因其非线性、自适应及学习特性而受到极大关注,并在诸多领域都取得成功的应用,如模式识别与图像处理、控制与优化、预测、通信等。 语音信号本质上是一个非平稳和非线性的过程,但一直以来,传统的语音处理方法都采用一种线性预测方法来处理,这就无法适应语音信号的非线性特征。而现有神经网络非线性滤波方法对矢量激励语音编码尚无有效的解决方案。 本课题首先针对线性滤波方法的不足,在语音编码系统的预测中引入神经网络模型,并研究了基于神经网络的语音编码系统的结构、适合于语音后向预测的神经网络的结构和学习算法,并且针对算法实时性的要求通过固定部分微变系数改进了BP网络训练过程缩短了训练时间,实验表明本算法比G721信噪比提高1.5-2dB。更多还原

【Abstract】 Artificial neural network (ANN) is complicated information processing one made of many processing units. This network has the ability of learning memory and input information trait extracting. Now it receives great attention and gets successful application such as mode recognize and image processing , control and optimize , predict, communication etc.Speech signal is got in essence non-stationary and nonlinear. But all along, traditional speech processing method uses linear prediction, it do not adapt well to the nonlinear characteristic of speech signals, and there is not any effective schemes about vector exciting speech coding algorithm in existing nonlinear prediction algorithm based on neural network.Aimed at this shortage about linear prediction, the non-linear predictors based on ANN are researched in this article. ANN is used to replace the conventional LP technology. At first, the structure of the system of speech coding based on ANN is researched. Secondly, the structure and the algorithm of the ANN fitted to speech signal are analyzed. Due to the complicated learning of ANN, it is difficult to implement speech coding in real-time. This article ameliorates the process of BP neural network training and shortens the time of BP neural network training by making part of coefficients of BP neural network fixedness. The experiment results indicate that the speech SNR based this arithmetic has increased 1.5-2 dB compared to that of G721 by CCITT.Vector quantization (VQ) is efficient method in speech coding algorithm, and there are not any effective schemes about vector exciting speech coding algorithm in existing nonlinear prediction algorithm based on neural network. This paper presented a new concept on nonlinear inverse filter based on BP neural network. A unit transform nonlinear filter with center tap can be got after off-line network training, which was divided into a positive filter and a inversefilter from the middle of tap; the speech passed through the positive filter and was formed into exciting vector; The exciting codebook can be obtained by training wave vector using LBG method. The coding end searches the codebook to product the optimal exciting vector. In decoding end it was nonlinear inverse filtered and the synthesis speech can be got. For shortening search time, this paper uses the search method based Fractal, trains the trained codebook again into some son-codebook and gets representative code of every son-codebook. When searching, at first gets representative code that is similar to originality exciting, then search relevant son-codebook. This search method shortens search time by two-quantity unit. Based on these theories, this article designs and develops 8 kbps low delay vector exciting speech coding algorithm based on by neural network. In this algorithm coefficients of BP neural network are fixed, so its MIPS is 42.2 in aspects of complication. The experiment has shown the SNR is 15.3323 in 30 sentences.更多还原