【作者】 王宏； 潘金贵；

【Author】 Wang Hong~(1,2) Pan JinGui~1 1.State Key Laboratory for Novel Software Technology,Nanjing University,Nanjing,210093; 2.Institute for Computer Application & Research,Changji University,Changji,831100

【机构】 南京大学软件新技术国家重点实验室；

【摘要】 在文本有关的说话人识别中,HMM 是目前常用的方法之一,而在文本无关的说话人识别领域,主流的方法则是 GMM。HMM 假设给定状态的特征矢量相互独立,GMM 假设所有的特征矢量相互独立．上述假设显然不大符合实际,因此,本文提出一种基于矩阵正态分布(MND)的文本有关说话人识别方法,该方法提取常用词的归一化特征矩阵作为说话人的识别特征．实验结果表明 MND 模型有较好的识别性能．更多还原

【Abstract】 HMM has become one of the most popular statistical tool for text-dependent speaker recognition,and GMM has proven extremely successful for text-independent speaker recognition.HMM assumes that parameter vectors are independent in a given state,GMM assumes that all parameter vectors are independent.It is obvious that the frame-independent assumption condition in both HMM and GMM does not accord with the speech signal,so,A new text-dependent speaker recognition method based on MND(Matrix Normal Distribution)is proposed in this paper,which takes the normalized parameter matrix extracted from some very frequent words as speaker feature. Experiment shows that MND has good speaker recognition performance.更多还原