基于子带处理多分类器融合的说话人识别  

Speaker recognition based on sub-band processing and multi-classifier combination

作　　者：崔国辉[1] 田岚[1] 

机构地区：[1]山东大学信息科学与工程学院,山东济南250100

出　　处：《山东大学学报（工学版）》2004年第4期108-111,共4页Journal of Shandong University（Engineering Science）

基　　金：山东省自然科学基金资助项目 (Y2 0 0 2G13 )

摘　　要：提出了一种基于子带处理多分类器融合的说话人识别方法 .宽带语音信号通过Mel滤波器组转变为多个子带信号 ,对各子带数据独立分析提取相应的特征参数 ,进而对每个子带分别建立识别模型进行判决 ,最后利用分类器融合规则 ,给出总体判决 .研究表明 ,该方法在子带数目选为 16时可以得到最好的识别效果 ,并且在有窄带噪声的情况下 ,子带多分类器融合法比宽带语音数据建模表现出更好的鲁棒性 .A method of speaker recognition based on sub-band processing and multi-classifier combination is presented. The wideband speech signal is filtered into several sub-bands through Mel filter bank,and feature extraction is individually performed on each sub-band. The resulting sequences of feature vector are modeled and identified on each sub-band,and then the outputs from each separate classifier are fused to produce an overall decision. The experiment results show that the sub-band processing approach has the best performance when the number of sub-bands is 16,and has a higher robustness for narrow-band noise compared with the method of wideband speech features modeling.

关 键 词：说话人识别 子带处理 分类器融合 MFCC 噪声 

分 类 号：TN912.34[电子电信—通信与信息系统]