Abstract: Aimming at the shortage of Sparse Kernel Principal Component Analysis（SKPCA） in feature extraction，a novel feature extraction method based on the kernel K-means clustering and the SKPCA for speaker recognition is proposed.Here kernel K-means clustering is to divide all the frames of each sample into a given amount of clusters，since the resulted clustering centers can represent better the clusters they belong to，the clustering is replaced by the clustering center，and the dimensions of kernel matrix are decreased accordingly.This method reduces storage and computational complexity，it guarantees to reduce data can represent the original data well and information loss is minimum.The experimental results show the proposed approach do not affect veracity，improve recognition rate，and meet the requirement of speaker recognition in terms of practicability.

Key words: Kernel Principal Component Analysis（KPCA）, sparse KPCA, kernel K-means clustering, speaker recognition

摘要： 针对稀疏核主成分分析方法在特征提取中的不足，提出了一种基于核K-均值聚类的稀疏核主成分分析（Sparse KPCA）的特征提取方法用于说话人识别。通过核K-均值聚类的方法对语音帧进行聚类，由于聚类的中心能够很好地代表类内的特征，用中心样本帧取代该类，减少了核矩阵的维数，然后再采用稀疏KPCA方法对核矩阵进行特征提取。该方法能够减少存储空间和计算的复杂度，它保证约简后的数据能够很好地代表原始数据并且在约简过程中信息损失最小。实验结果验证了提出的方法在不影响识别率的前提下提高了识别速度，满足了说话人识别的实用性要求。

关键词: 核主成分分析（KPCA）, 稀疏KPCA, 核K-均值聚类, 说话人识别