自适应高斯混合模型及说话人识别应用  

Adaptive Gaussian Mixture Model and Its Application in Speaker Recognition

作　　者：王韵琪 俞一彪[1] 

机构地区：[1]苏州大学电子信息学院,江苏苏州215006

出　　处：《通信技术》2014年第7期738-743,共6页Communications Technology

基　　金：国家自然科学基金(No.61271360);苏州市应用基础研究计划(No.SYG201230)~~

摘　　要：高斯混合模型采用固定混合数结构的建模方法并不符合说话人语音特征分布的多样性,从而出现过拟合或者欠拟合的情况并影响系统的识别性能。提出一种混合数可变的自适应高斯混合模型并将其应用于说话人识别。模型训练中根据说话人语音特征参数分布的聚类特性,采用吸收合并与分裂机制动态调整混合数以获得更加精确的拟合性能,提高系统识别率。实验结果显示,在特征参数MFCC和BFCC(Bilinear Frequency Cepstrum Coefficients)下相对误识率分别下降了41.41%和22.21%。Gaussian mixture model（ GMM） is not reasonable to train the model with a fixed mixture number according to the diversity of speaker＇s acoustic feature distributions,and may leads to over-fitting or under-fitting then to drop system performance. This paper proposes an adaptive Gaussian mixture model（ AGMM） with variable component numbers and applies to speaker recognition. With the cluster property of speaker＇s actual acoustic feature distribution,an absorb-merge-split mechanism is adopted to adjust the component number dynamically in the model training. The experimental result indicates the relative false recognition rates decline by 41. 41% and 22. 21% with parameter MFCC and BFCC respectively.

关 键 词：说话人识别 自适应高斯混合模型 双线性频率倒谱系数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]