Abstract: Channel mismatch of speaker verification system results from mismatches between training and test voices. A speaker verification system is proposed which based on Nuisance Attribute Projection(NAP). Voices from a large number of known channel information are used to train higher-dimensional space mapping matrix. Then, Gaussian Mixture Model(GMM)-supervectors which are used as Support Vector Machine(SVM) inputting parameter are projected in order to eliminate the effects of channel information. Experimental result indicates that the system can reduce the negative effects of channel mismatch.

Key words: Support Vector Machine(SVM), Gaussian Mixture Model(GMM)-supervetor, mismatch compensation, mapping matrix, Nuisance Attribute Projection(NAP)

摘要： 在说话人确认系统中，由于训练和测试语音来自不同的信道，会产生信道失配现象。为此，提出一种基于扰动属性投影的说话人确认系统。利用有信道标注信息的语音训练出高维空间映射矩阵，通过映射将自适应得到的超矢量作为SVM的输入，削弱说话人特征中的信道信息。实验结果表明，该系统能降低信道失配带来的负面影响。

关键词: 支持向量机, 高斯混合模型超矢量, 失配补偿, 映射矩阵, 扰动属性投影

CLC Number: 

• TN912.3