基于模糊核熵的短语音说话人识别  

Speaker Recognition with Little Data Based on Fuzzy Kernel Entropy

作　　者：林琳[1] 王树勋[1] 陈建[1] 

机构地区：[1]吉林大学通信工程学院,吉林长春130022

出　　处：《系统仿真学报》2008年第16期4368-4372,4378,共6页Journal of System Simulation

摘　　要：为了解决由于模糊加权指数选取不当而导致系统性能下降的问题,将熵函数引入到核映射的特征空间中,提出了一种基于模糊核熵的短语音说话人识别新方法。通过定义特征空间中的模糊核熵目标函数,设计了模糊核熵矢量量化器,在高维特征空间中对说话人的语音进行训练和识别匹配。引入熵的算法具有更加清晰简洁的数学形式和物理含义。为了进一步提高系统性能,利用模拟退火法对熵函数中的模糊熵度进行全局优化,提出了一种基于模拟退火的模糊熵度更新方法。实验表明,该方法对于较短的训练语音,其识别效果优于高斯混合模型和模糊核矢量量化。In order to resolve the problem that the system performance decreased due to the unsuitable fuzzy weighted exponent, using the entropy function in the kernel mapping feature space, a novel speaker recognition method based on fuzzy kernel entropy was p^gposed. By defining the fuzzy kernel entropy objection function in the feature space, fuzzy kernel entropy vector quantization was designed to train speakers＇models and make identification decision in high-dimensional feature space. The algorithm with entropy function has more beautiful form and the clearer physical meaning. To improve the system performance, a fuzzy entropy degree updating method with simulated annealing was proposed, which could optimize the fuzzy entropy degree globally. Experimental results show that it can obtain better results than the Gaussian mixture model and fuzzy kernel vector quantization method in the case of the little training data.

关 键 词：熵函数 模糊核熵矢量量化 模拟退火 说话人识别 短语音 

分 类 号：TN912.3[电子电信—通信与信息系统]