摘要
语音特征提取问题取决于参数。针对特征参数识别准确性,通过核主成分分析方法应用于语音特征提取中。但核主成分分析方法的计算过于复杂,不利于提高系统实时性。为提高语音识别系统的鲁棒性和增强实时性,提出基于K-均值聚类的核主成分分析方法。通过K-均值聚类的方法对每个语音信号的语音帧进行聚类,采用聚类的中心代表类的特征,再用核主成分分析方法进行特征提取,不但减少了存储空间和计算的复杂度,而且通过把原始特征向量向低维子空间投影,达到降噪和去冗余的效果。仿真结果证明:所提方法在相似识别率的情况下提高了识别速度,能满足语音识别的实时性要求,并在噪声环境下具有较好的鲁棒性。
Kernel principal component analysis method has been applied widely in speech feature extraction at present,and it is helpful for improving robust of speech recognition system,but its computation is too complicated,which is the disadvantage of enhancing system real-time performance.Focusing on the computation drawbacks of the method,a new kernel principal component analysis method based on K-means clustering is proposed.In the new method,all the frames of speech signal is divide into a given amount of clusters by K-means clustering,the clustering center represent features of the clusters,and then the features were extracted by kernel principal component analysis.The new method not only reduces storage and computational complexity,but also maps the original eigenvector to the low dimension space which has effects of de-noise and eliminating tedious information.Simulations show the proposed method improves recognition speed and meets the needs of real-time of speech recognition at the similar recognition rate,and it has better robust under noisy environment.
出处
《计算机仿真》
CSCD
北大核心
2011年第6期393-396,共4页
Computer Simulation
关键词
核主成分分析
特征提取
语音识别
Kernel principal component analysis
Feature extraction
Speech recognition
