A Novel Hierarichical Speaker Identification Method

This paper proposes a novel hierarchical speaker identification method to save the speaker identification and training time, viz. First is to get a coarse decision by a fast scan all registered speakers using PCA classifier to found M possible target speakers; then is to get a final decision by the proposed Multi-Reduced Support Vector Machine (MRSVM). And the MRSVM has two reduction steps to reduce training time and the memory size for SVM. Firstly, speech feature dimensions are reduced by using PCA transform, the noise is removed from speech simultaneity; secondly, the training data are selected at boundary of each cluster as Support Vectors (SVs) by using Kernel-based fuzzy clustering technique. The experiment results show that the training data, time and storage size can be reduced remarkably by using the proposed reduction method, and the identification velocity is improved greatly by the hierarchical identification method and the system has better robustness.