针对一些说话人识别方法在应用中要面对海量的集外数据,其很低的错误接受率也会带来大量识别错误的情况,进行了极低错误接受率的说话人识别技术的研究,以求在保证召回率的前提下,将错误接受率降低至约万分之一的水平.研究的重点是对经典的高斯混合模型-通用背景模型(GMM-UBM)方法进行了改进,加入一个确认判决机制来进一步拒绝集外误识,尝试了三种确认方法--基于粗粒度分析窗的方法、基于集外竞争模型的方法、基于变化状态统计矢量的方法.实验结果表明,这三种方法都能够有效降低错误接受率指标,其中基于变化状态统计矢量的方法取得了最好的效果.%Seeing that some speaker recognition methods have to deal with massive non-target speech data when in application and a rather low false acceptance rate of them can cause a large number of recognition errors, the authors of the paper carried out the research to develop a speaker recognition technique with the very low false acceptance rate of less than 0.0001 under the condition of guaranteeing a reasonable recall rate. The key of the work was to improve the Gauss mixture model (GMM)-uniform background model (UBM) based method by adding a verification module to further avoid false recognition of non-target data. Three verification methods, respectively based on coarse-grained analysis window,non-target competing model, and statistic vector of change status, were investigated. The experimental results showed that they all reduced false alarms effectively and the best performance was given by the method based on statistic vector of change status.
展开▼
