基于子聚类约简支持向量机的说话人识别研究

Speaker identification based on sub-clusters reduced SVM

由于支持向量具有边界性,在利用语音训练集对基于支持向量机(SVM)的说话人识别系统进行训练之前,需要对该训练集进行约简。考虑到该训练集一般十分庞大且具有非线性可分的特性,提出子聚类约简的概念。首先对训练集进行模糊核子聚类并过滤掉非边界的聚类区,然后依照提出的算法对保留的聚类区中的向量集做进一步地约简,使支持向量集更加集中在边界。理论和实践表明,经过两层的约简既保留了充足支持向量,保证了SVM良好的泛化性能,又提高了系统的时间和空间效率。

Because supporting vectors are near to the border,before training speaker identification system based on SVM using speech training assemble,it is needed to reduce the training assemble.Because the training assemble are tremendous and non-linear,a concept called sub-clusters reduced SVM is presented.Firstly,this paper uses kernel-based fuzzy sub-clusters to filter the non-border cluster areas,then cuts the redundant vectors within the remaining cluster areas.Theory and experimental results show that the time and storage efficiency can be enhanced remarkably by using the training and identifying algorithm,and the system has better robustness.