语音转换中特征参数及其转换方法的研究

发布时间：2018-10-21 08:59

【摘要】：语音转换技术实现的是两个不同说话人声音个性特征之间的转变,即以目标说话人的声音特征为导向,通过改变源说话人的个性特征使之具有目标说话人的特征,从而改变源说话人话音的听觉效果。本文主要研究语音个性特征参数的转换方法,实现语音参数的有效利用以及转换效果的增强。主要工作如下:(1)对基音频率、语速等韵律特征的调整进行研究。本文在实现基音频率转换的同时,提出用高斯法作为源和目标语音时长的映射规则,并利用按照获得的时长比例来进行内插的方法实现目标语音时长的调整,从而使合成的语音时长更接近目标语音,提高语速的相似度。(2)对声道特征参数的转换规则进行研究。人工神经网络的泛化属性有助于实现说话人特征的转换,但网络的训练通常存在隐含层节点较多导致网络结构较复杂的问题。因此,本文提出基于改进的径向基函数神经网络的声道特征参数转换方法。该方法用K均值算法作为神经网络中心值算法,并利用粒子群优化算法对网络隐含层的节点数进行优化,有效地提高了径向基函数网络对多维非线性特征参数拟合与转换效率,从而提高转换语音与目标语音的相似度。(3)对语音转换系统进行完善。由于提取的声道特征参数通常具有数据量较大的特点,而且对于同一条将要转换的测试语音,不同数据段训练得到的转换规则其转换效果有所不同。为了充分利用提取的参数,本文提出用高斯混合模型对提取的大量特征参数进行预处理,得到数据量较小并且特征性较强的训练参数。因此,在改进神经网络转换算法的基础上,加入特征参数预处理模块以及时长调整模块,完善语音转换系统的同时提高转换语音的质量。
[Abstract]:Speech conversion technology realizes the transformation between two different speakers' voice personality characteristics, that is, the target speaker's voice characteristics are oriented, and the source speaker's personality characteristics are changed to make it have the target speaker's characteristics. Thus, the auditory effect of the source speaker's voice is changed. In this paper, the conversion method of speech personality parameters is studied to realize the effective use of speech parameters and the enhancement of the conversion effects. The main work is as follows: (1) the adjustment of pitch frequency, speech speed and other prosodic characteristics is studied. In this paper, while realizing pitch frequency conversion, Gao Si method is proposed as the mapping rule of source and target speech time, and the target speech duration adjustment is realized by interpolation method according to the obtained timespan ratio. Thus, the synthesized speech time is closer to the target speech, and the similarity of speech speed is improved. (2) the conversion rules of vocal channel feature parameters are studied. The generalization attribute of artificial neural network is helpful to realize the transformation of speaker's feature, but the training of the network usually has the problem that there are many hidden layer nodes and the network structure is more complicated. Therefore, this paper presents a method of acoustic channel characteristic parameter conversion based on improved radial basis function neural network. In this method, K-means algorithm is used as the center value algorithm of neural network, and particle swarm optimization algorithm is used to optimize the number of nodes in the hidden layer of the network, which effectively improves the efficiency of fitting and converting the multi-dimensional nonlinear characteristic parameters of the radial basis function network. In order to improve the similarity between speech and target speech. (3) improve the speech conversion system. Because the extracted feature parameters usually have the characteristics of large amount of data, and for the same test speech to be converted, the conversion rules obtained by different data segment training have different conversion effects. In order to make full use of the extracted parameters, this paper proposes to preprocess a large number of feature parameters by Gao Si mixed model, and obtain the training parameters with small amount of data and strong characteristics. Therefore, on the basis of the improved neural network conversion algorithm, the feature parameter preprocessing module and the time length adjustment module are added to improve the speech conversion system and improve the quality of the converted speech.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN912.3

【相似文献】