PCA变换下的GMM-SVM话者确认研究

New Combination Method of GMM-SVM Based on PCA for Speaker Verification

针对支持向量机(SVM)输入参数不能充分利用高斯混合模型(GMM)均值、方差、权重所携带的说话人信息,而导致与文本无关话者确认系统性能下降的问题,本文结合GMM的均值、方差、权重,提出一种新的、基于自适应后GMM的,SVM模型输入特征提取方法。在NIST 06语音数据库上的实验表明,本方法将等误识率(EER)从高斯混合模型-通用背景模型(GMMUBM)系统的8.49%,下降到基于离散余弦变换(DCT)变换GMM-SVM系统的4.16%,以及基于主元成分分析(PCA)GMMSVM系统的3.3%.

Concerning the ineffectiveness of Support Vector Machine （ SVM） input parameters in taking full advantage of speaker in- formation carded by mean, variance, and weight of Gaussian Mixture Model （GMM ）, which lead to the speaker verification perform- ance degradation problem. In this paper, a new kind of approach is proposed to extract feature for Support Vector Machine （ SVM ） from adapted Gaussian Mixture Model （ GMM） in text-independent speaker verification system. In the NIST06 speech databases, ex- perimental results show that this method reduce the Equal Error Rate （EER） from 8.49% of GMM-UBM system,down to 4. 16% of the DCT-based GMM-SVM system and 3.3% of the PCA-based GMM-SVM system.