基于深层说话人矢量的说话人检索

Speaker retrieval based on deep speaker vector

为了解决浅层特征不能有效刻画说话人特征,导致说话人检索率不高的问题,提出了一种基于深层说话人矢量的说话人检索方法.使用受限波尔兹曼机逐层构建一个多层的深层特征提取器用以提取说话人深层特征.为说话人构建基于深层特征的深层说话人矢量.通过计算要检索的说话人的深层说话人矢量和检索库中的说话人深层特征之间的最小距离,对目标说话人进行检索.实验结果表明:在深层特征下,使用深层说话人矢量可以检索到绝大部分的目标说话人;随着深度层数的增加,检索率先增后减,检索率最高对应的深度层数是7;随着深度层数的增加,检索时间非线性增加.

In order to solve the problem that shallow feature can not depict speakers effectively,spearker reterieval rate is low,a method of speaker retrieval was proposed based on deep speaker vectors.Firstly,a multi layers deep feature extractor was constructed by using restriced boltzmann machines（RBM）training layer by layer to extract speaker deep feature.Secondly,deep speaker vectors were built.Lastly,object speaker was retrieved by calculating the minimal distance between deep speaker vectors of retrieval speaker and deep feature of speakers in retrieval library.Experimental results demonstrate that under deep feature,most of speakers can be retrieved using deep speaker vectors.Retrieval rate of the first and second layer are lower than mel-frequency cepstral coefficial（MFCC）and the third layer is the same as MFCC.Retrieval rate increases firstly and decreases later with the increasing of the depth of layers,and the highest retrieval rate corresponding to depth layers is 7.Retrieval time increases non-linearly with deep layer increasing.