【摘要】 在与文本无关的说话人确认研究中,特征映射是减少信道影响的最有效的方法之一。但是随着信道的复杂化,特征映射的效果很难体现出来。本文在混合高斯模型的均值超矢量上采用EM算法对信道类型进行聚类,然后通过特征映射的方法在特征参数域中减去信道的影响。在NIST SRE 2006数据库1conv4w-1conv4w任务上,采用本文方法的系统相对基线系统在等错误率上降低了18%。更多还原

【Abstract】 In the text-independent speaker verification research,feature mapping is one of the most effective methods to reduce the channel bias.In this paper,the channel information is clustered by EM algorithm in the GMM(Gaussian mixture model) supervector space,then the channel bias is subtracted from the speech feature by feature mapping.In the NIST SRE(speaker recognition evaluation) 2006 1conv4w-1conv4w corpus,EER(equal error rate) of the proposed system can be reduced by 18% against the baseline GMM system.更多还原