摘要
大多数说话人确认系统都设置一个背景模型用于描述假冒者的特性。文中提出一种新的说话人确认背景模型,对所有说话人采用同一全局背景模型(UBM),并为每个说话人建立一个竞争者模型(cohortmodel)和一个疏远者模型(c-cohortmodel)。在全局背景模型不能做出准确判断的情况下,启用竞争者模型或疏远者模型再次进行判决。该模型充分利用了相近者模型和疏远者模型的特性。实验表明新的背景模型使系统性能有明显的提高。
Most speaker verification systems use background model for describing imposters . Based on analysis of general background models, a new model combining the universal background model （UBM） with a cohort model and a c-cohort model is proposed. In the proposed speaker verification system, the same UBM is set for all speakers, a cohort model and a c-cohort model is set for each speaker. When the UBM fails to give a definite decision, the cohort model and c-cohort model is used to make a verdict. The new background speaker model takes advantages of using UBM or the cohort model or c-cohort model alone. Experimental resuits show that it has an error rate lower than conventional technique＇s.
出处
《信息化研究》
2010年第2期19-23,共5页
INFORMATIZATION RESEARCH
关键词
说话人确认
背景模型
全局背景模型
竞争者模型
疏远者模型
speaker verification
background model
universal background model
cohort model
c-cohort model
