摘要
为提高说话人识别的性能,提出将CCA与PCA联合用于说话人特征降维的方法:先用CCA融合基于声道模型的LPC特征和基于听觉模型的MFCC特征,提升这两类不同特征的相关性;然后用PCA进一步去除冗余特征,降低有效特征的维数。实验显示,这两种降维方法联合的降维效果与单一的CCA降维、PCA降维或手动降维的效果比有明显提高。
With the purpose of improving the performance of speaker recognition, a method of dimension reduction in speaker＇ s characteristics by jointing CCA and PCA is proposed. Firstly, LPC characteristics based on acoustic models and MFCC characteristics based on auditory models are blended by CCA method so as to enhance the eorrelativity between LPC and MFCC. After that the PCA method is used to eliminate redundant characteristics so as to reduce the effective characteristic dimensions of speech signal. Experiments show that the efficiency of dimension reduction of this novel method that joints CCA and PCA is significantly improved comparing to that of traditional methods while only using CCA dimension reduction, PCA dimension reduction or manual dimension reduction.
出处
《计算机与现代化》
2013年第6期16-19,共4页
Computer and Modernization
基金
江苏省自然科学基金资助项目(BK2009059)
解放军理工大学预研基金资助项目(2009TX08)
关键词
说话人识别
典型相关分析
主成分分析
高斯混合模型
特征降维
线性预测系数
美尔频率倒谱系数
speaker recognition
canonical correlation analysis （CCA）
principal components analysis （PCA）
Gaussian mixture model （GMM）
dimensional reduction
linear prediction coefficient （LPC）
Mel frequency cepstrum coefficient（MFCC）
