论文部分内容阅读
对于电话手机语音的文本无关话者确认,运用联合因子分析构建话者信息子空间与信道信息子空间来进行失配信道补偿取得了较好的效果.然而研究表明,信道信息子空间仍然包含了可以用来区分话者的信息.因此,本文运用一种既包含话者信息又包含信道信息的全变量信息子空间来提取i-vectors低维特征矢量,再运用类内协方差规整进行失配信道补偿,最后用补偿后的i-vectors特征矢量构建支持向量机话者模型.在NIST08数据库上实验表明,本文所构建系统的性能在等误识率和最小检测代价函数上有相对近70%的提高.
Confirming the text irrelevant speaker of the telephone handset voice and using the combination factor analysis to construct the speaker information subspace and the channel information subspace to obtain the mismatch channel compensation has achieved good results. However, the research shows that the channel information subspace still contains Can be used to distinguish the information of the speaker.Therefore, in this paper, an all-variable information subspace containing both speaker information and channel information is used to extract i-vectors low-dimensional eigenvectors, and intra-class covariance Channel compensation, and finally use the compensated i-vectors feature vector to build SVM speaker model.Experiments on NIST08 database show that the performance of the system constructed in this paper has a relative 70% Improve.