摘要
从线性预测(LP)残差信号中提出了一种新的特征提取方法,这种特征跟单个的说话人的声道密切相关。通过把HAAR小波变换运用于LP残差而获得了一个新的特征(HOCOR)。为了进一步提高系统的鲁棒性和辨识率,在采用分级说话人辨识的基础上,将基音周期的高斯概率密度对GMM分类器的似然度进行加权,形成新的似然度进行说话人辨识。试验结果显示,所提出系统的鲁棒性和辨识率都有所提高。
A novel feature extraction method from LP residue signal is proposed.This kind of feature has good relation with vocal tract of speaker.A novel feature（HOCOR） is acquired by applying LP residue with HAAR.In order to improve the robustness and identification rate of the system,hierarchical speaker identification is proposed.Then the likeliness of GMM classifier is weighted by the Gauss probability density of the pitch to form the novel likeliness which is used for speaker identification.The experiment result shows that the robustness and identification rate of the system proposed are both improved.
出处
《计算机工程与应用》
CSCD
北大核心
2010年第11期122-124,共3页
Computer Engineering and Applications
