基于非参数直方图模型的鲁棒说话人识别算法  

Novel Non-parametric Model for Robust Speaker Recognition

作　　者：李燕萍[1] 唐振民[1] 丁辉[1,2] 张燕[1,3] 

机构地区：[1]南京理工大学计算机科学与技术学院,南京210094 [2]嘉兴学院数学与信息工程学院,嘉兴314001 [3]金陵科技学院信息技术学院,南京210006

出　　处：《数据采集与处理》2010年第1期81-85,共5页Journal of Data Acquisition and Processing

基　　金：浙江省教育厅科研基金(Y200805349)资助项目;浙江省自然科学基金(Y1090649)资助项目

摘　　要：建立一种非参数模型来刻画说话人的特征分布,并采用地面移动距离来度量分布之间的相似性。该方法能有效地利用有限的数据表达说话人的身份信息,直接计算特征分布与测试语音分布之间的距离,与传统的矢量量化和高斯混合模型相比,不需要通过对所有语音帧计算总平均失真误差和最小相似度,计算简单,主要能够降低系统对数据量的依赖性。并且通过自适应直方图均衡化方法对原始语音特征进行修正,使得噪声环境下获得的语音特征经过修正后更符合真实分布,增强了特征的抗噪性。实验表明,本文提出的方法在噪声环境下的短语音说话人识别系统中表现出较强的优势。The limitation of training speech data and the noise disturbance are two key factors influencing the praetieality of the speaker reeognition system. Firstly, this paper establishes a novel histogram nonparametrie model to depict the feature distribution of each speaker and adoptes earth mover distance as the similarity distance, thus it can efficiently use finite speech data to depict the speaker identification information, compared the distribution of the speaker model with the distribution of the testing feature vectors. Secondly, the feature distributions by using adaptive histogram equalization are modified. It can reduce the noise influence and strengthen the robustness of features compared with the conventional histogram equalization. The theory analysis and experimental results demonstrate that the method has the advantages of simple calculation, reducing the amount of data and improving the recognition performanee.

关 键 词：说话人识别 非参数模型 地面移动距离 自适应直方图均衡化 

分 类 号：TP912.34[自动化与计算机技术]