基于树形通用背景模型的高效说话人辨认

为了提高基于G auss混合模型通用背景模型(GMM-U BM)的说话人辨认系统的运算效率,提出一种基于树的核心挑选算法(TBK S),通过将U BM中的各个G auss分布按组织成树形结构,来减少从中挑选核心分布的运算量。实验结果表明:对1 000个说话人进行辨认,TBK S与现有的基于特征矢量重排序的剪枝算法(ORBP)相结合,将基于GMM-U BM的辨认系统的运算速度提高21.9倍,误识率却只上升不到4%;TBK S和ORBP相结合,可大幅度提高GMM-U BM系统的运算效率,而基本不降低识别率。

Tree-structure universal background model-based efficient speaker identification

A tree-based kernel selection(TBKS) algorithm,in which all the Gaussian components in the universal background model are clustered hierarchically into a tree structure for efficient kernel selection,was developed as a computationally efficient approach for Gaussian mixture model-universal background model-based speaker identification.In tests on a database of 1 000 speakers, integration of the TBKS algorithm and an observation reordering-based pruning(ORBP) method improved the computation speed by a factor of 21.9 with only 4% increase in error rate compared with the baseline GMM-UBM system.The experimental results show that by integrating the TBKS and ORBP algorithms,the computation efficiency of the GMM-UBM system can be significantly improved with almost no reduction in recognition rate.