【作者】 杨延龙；

【导师】 高西全；

【作者基本信息】 西安电子科技大学 ， 通信与信息系统， 2010， 硕士

【摘要】 本文主要对与文本无关的说话人识别的基本原理和相关算法进行了深入的分析和研究。在端点检测中,对几种常用的检测算法进行了分析和研究,针对单个参数用于端点检测所存在的问题,本文将短时能量和Renyi熵的结合能量-Renyi熵(ERE)参数用于端点检测,取得了较好的效果。在语音增强中,重点分析和研究了β?order自适应谱减法。在此基础上,本文采用谱增益迭代的方式来进一步逼近真实语音谱,并将该算法应用到近几年发展起来的CMSBS参数提取过程中的输出子带能量增强,取得了很好的效果。在识别模型方面,在对GMM基本原理和相关算法分析的基础上,重点研究了GMM参数估计算法-SGML算法。该算法通过自分裂的方式寻找最佳的模型混合度,而且能够在分裂过程中不断提高参数估计精度,很好地解决了使用EM算法估计参数时面临的模型混合度较难选取和对初始值较为敏感的问题。通过对上述算法的深入分析,在VC平台上对其性能进行了验证。更多还原

【Abstract】 This paper mainly focuses on the study of the basic principles and related algorithms of text-independent speaker recognition. In the endpoint detection, the research is focused on the endpoint detection algorithm. With the shortages of a single parameter which is used for endpoint detection, the combination of energy and Renyi entropy: energy- Renyi entropy parameter is used for endpoint detection. In the speech enhancement, the research is mainly focused on the adaptiveβ?order spectral subtraction. On the basis, the adaptiveβ?orderspectral subtraction based on the iteration of spectral gain function is used to enhance speech spectrum. This algorithm is also applied to the enhancement of the output sub-band energy of CMSBS parameter extraction. Extensive experiments indicate the efficiency of the algorithm. In recognition model, with the study of principles and related algorithms, the research is mainly focused on the SGML algorithm which is used for GMM parameter estimation. The SGML algorithm not only searches the mixed-degree of GMM with self-splitting method but also improves the parameter estimation accuracy on the process of splitting. It is a good solution to the problems that the mixed-degree is difficult decided and EM algorithm is sensitive to the initial value. Furthermore, the performance of the location algorithms is verified on VC platform.更多还原