摘要
文中研究表明 ,反映说话人特征信息的特征参数矢量的各个分量通常具有不同的分布 ,对正确识别说话人身份的有效性是有差别的。文中将这种有效性差别作为权重矢量反映到失真测度计算公式中 ,提出了一种新的失真测度 ,即方差归一化失真测度。该失真测度可有效提高话者识别系统的识别性能。进一步的实验还表明 ,该失真测度能提高话者识别系统的时间鲁棒性。文中同时还给出了适合于话者识别的参数归正方法 :帧内幅度归正。
The research shows that each dimension of the feature vector has different distributions and different validities for speaker recognition. This paper presents a new distortion measurement, i.e. normalized square mean distortion measure, which can combine the validities of the different distributions with the calculation of the distortion. Experiments demonstrate that the distortion measurement presented here can not only improve the speaker recognition system performance, but also enhance its time robustness. Additionally, a good normalization method, that is an amplitude normalization within frames, is also presented.
出处
《数据采集与处理》
CSCD
2000年第2期157-161,共5页
Journal of Data Acquisition and Processing
基金
国家自然科学基金
关键词
失真测度
鲁棒性
VQ话者模型
话者识别
distortion measurement
robustness
normalized square mean
speaker model
vector quantization