基于鉴别性i-vector局部距离保持映射的说话人识别

Speaker recognition based on discriminant i-vector local distance preserving projection

为了进一步提高i-vector说话人识别系统的性能,该文提出了一种鉴别性i-vector局部距离保持映射(discriminant i-vector local distance preserving projection,DIVLDPP)的流形学习算法。该算法以i-vector间的Euclid距离作为度量准则,并以最小化同类点间距离同时最大化异类近邻点间距离的鉴别性准则作为优化目标函数,利用求解广义特征值的方法,得到最终的投影映射矩阵。在美国国家标准技术局2008年说话人识别核心数据集上的实验结果表明:该算法可以明显提高目前i-vector说话人识别系统的性能。

The performance of the popular i-vector based speaker recognition system is improved by a manifold learning algorithm named discriminant i-vector local distance preserving projection（DIVLDPP）.This algorithm uses the Euclidean distance to measure the i-vector space.The target function minimizes the distance between the same speaker samples and maximizes the distance between neighbouring samples of different speakers.A linear mapping matrix is obtained by solving a generalized eigenvalue problem.Tests on the speaker recognition evaluation data corpus released by the US National Institute of Standards and Technology in 2008 demonstrate that this i-vector system provides better speaker recognition performance.