【摘要】 身份-矢量(identity-vector, i-vector)方法作为说话人确认领域中的主流方法之一,能够通过学习总变化空间来获取有效的低维说话人特征——i-vector特征.但是当开发集数据不充足时,会导致学习到的总变化空间模型误差较大;同时,还无法有效确认此时的总变化空间是否因为预先设置的维度过高而学到了冗余信息.为此,本文将贝叶斯主成分分析(Bayesian Principal Component Analysis, BPCA)引入总变化空间的学习过程中,利用其来为总变化空间引入更多的先验信息,从而对开发集数据中包含的信息进行补充,并在先验信息的约束下削弱总变化空间中无效维的影响.实验结果表明,当开发集数据不充足时,相比于传统的总变化空间学习方法,BPCA方法能够有效提升说话人确认系统的识别性能.更多还原

【Abstract】 As one of the most important methods in speaker verification, the identity-vector(i-vector) approach can obtain effective low-dimensional i-vector by learning the total variability space(TVS). However, when there is no sufficient development data, it will lead to a large error in the learned TVS model. Meanwhile, it is difficult to determine whether there is redundancy in the learned TVS due to the high preset dimension. To solve the above problems, the Bayesian principal component analysis(BPCA) is introduced into the learning of the TVS. And this proposed method can introduce more prior information into the TVS to supply more information. Additionally, under the constraint of prior information, the influence of invalid dimension in the TVS can be weakened. The experimental results show that when the development data is insufficient, the BPCA method can effectively improve the performance compared with the traditional TVS learning methods.更多还原