改进批处理RPEM算法用于说话人识别

针对传统EM算法训练GMM不能充分利用训练数据所属高斯分量信息, 从而在一定程度上影响说话人识别性能的缺陷, 采用RPEM (竞争惩罚EM)算法训练GMM, 并引入批处理RPEM算法解决RPEM算法运算量大、收敛速度慢的问题, 同时针对RPEM和批处理RPEM算法训练时方差优化存在的问题进行了改进, 提出了改进的批处理RPEM算法。在Chains 说话人识别数据库上的实验表明, 改进的批处理RPEM算法取得了相对于传统EM、RPEM以及批处理RPEM算法更好的性能, 还极大地提高了训练效率, 减小了运算量, 说明了提出的改进批处理RPEM算法用于说话人识别时的有效性。

Improved batch RPEM algorithm for speaker recognition

When the traditional EM algorithm was used to train GMM, it often failed to make full use of the information that the training data belongs to which Gaussian component. And it would influence the performance of speaker recognition to some extent. To solve this problem, this paper adopted the RPEM algorithm to train GMM. But the RPEM algorithm needed a large amount of computation, and its convergence speed was slow. So it introduced the batch RPEM algorithm to overcome the RPEM algorithm's above two shortcomings. However, there were also some problems with the optimization of the variance when it used RPEM or batch RPEM algorithm to train GMM. And it put forward the improved batch RPEM algorithm to solve these problems. The experiments that based on the chains speaker recognition database show that the improved batch RPEM algorithm not only achieves better recognition performance than other three algorithm's recognition performance, but also improves the training efficiency and reduces the amount of computation of the RPEM and batch RPEM.