语音转换中音段特征的建模和转换研究.pdf

南京邮电大学硕士研究生学位论文摘要 I 摘要语音转换就是将一个说话人(源说话人)语音中的个性特征信息进行转换,使之具有另一个说话人(目标说话人)的个性特征,从而使得转换后的语音听起来就像是目标说话人的声音的一种语音信号处理技术。该技术不仅具有重要的理论研究意义,而且具有良好的应用价值,它的研究及发展研究愈来愈受到国内外学者的关注。本文的主要工作和创新如下: (1)简要介绍了语音转换的一些应用价值和当前的主要经典算法,讨论了常用的语音个性特征参数,以及语音转换系统的基本原理。(2)研究了经典的基音频率转换方法,针对经典算法在不同程度上存在转换精度和合成语音质量不高的情况。本文提出基于STRAIGHT模型和BP神经网络的基音频率转换算法。客观测试和主观测试上都取得了较好的效果。(3)研究了基于神经网络的频谱包络转换方法,考虑到神经网络训练算法有很多,但大都有一定自身的缺陷,针对梯度下降法训练速度慢和易导致陷入局部最优的问题,引出了基于量子粒子群优化BP神经网络的算法,并将其算法应用到的语音谱包络转换中,由粒子群优化算法训练的BP神经网络捕获说话人的语音频谱包络映射关系,以实现不同说话人之间声音特性的转换,该方法在一定程度上提高了转换语音性能。本文在MATLAB平台上仿真,从主观和客观两个方面评价系统的性能。仿真结果表明, 本文所采用的转换方法能够取得较好的效果。关键词:语音转换,人工神经网络模型,量子粒子群优化算法,频谱包络转换,基音频率转换南京邮电大学硕士研究生学位论文 ABSTRACT II ABSTRACT Voice conversion is a voice signal processing technology that aims to transform the voiceof a speaker (source speaker),for it to be perceived by listeners as if it had been uttered by another speaker (target speaker),while keeping the semantics and emotional information unchanged. so that the source speaker’s voice sounds like the target speaker’s technology not only has the important theory research value,but also has mercial application value,and the scholars pay more and more attention to the current and trend researchat home and abroad. The main work and contributions are described as follows: Firstly, the paper briefly introduces the application valuation and classical algorithms,and discusses the familiar identity parameters of voice and some basic principles of voice conversion. Secondly, this paper researches some classical pitch frequency transformation algorithm. The classical algorithms exist lowness of transformation precision and ropiness of synthesized speech. For this reason, the pitch frequency transformation algorithm, based on the STRAIGHT+BP Neural Network, is proposed. The newmethod isevaluated by means of both objective and subjective tests, the experimentaion result has proved the validity of the method. Finally, this paper proposes