【作者】 刘巍巍； 张卫强； 刘加；

【Author】 LIU Weiwei ZHANG Wei-Qinag LIU Jia Tsinghua National Laboratory for Information Science and Technology Department of Electronic Engineering,Tsinghua University,Beijing 100084,China

【机构】 清华大学电子工程系,清华信息科学与技术国家实验室(筹)；

【摘要】 传统语种识别中训练数据库的规模庞大,对于语种分类有鉴别性的信息有大量重叠,且训练数据的不同信道条件、不同来源都会对训练和测试有一定干扰。针对这些问题,提出一种鉴别性向量空间模型(Discriminative-Vector Space Models,D-VSMs)建模方法。D-VSMs能够自动过滤训练集中信息重叠的数据,使得每一个支持向量机的训练数据都很有针对性,用很少的训练数据能取得很好的分类效果。在美国国家标准技术局2009年语种识别测试(NISTLRE2009)中,D-VSMs只用了原测试数据的30%参与训练,计算量是传统平行音素识别器后接向量空间模型(PPRVSM)的10%,等错率在30s、10s和3s的测试条件下分别比传统PPRVSM下降了12.75%、15.89%以及7.33%,是一种有效的语种识别技术。更多还原

【Abstract】 Conventional language recognition tasks are suffered from the large scale of the training data,in which most discriminative information is overlapped.Moreover,the non-language variability(eg.channel,speaker,and so on) also affects the performance of the language recognition systems.In this paper,we propose a method using discriminative vector space models(D-VSMs).Using this method,the overlapped training information are eliminated automatically.Thus every VSM is trained for one situation specially,and the whole system achieves a good performance.Compared with baseline,D-VSMs only use 30% training data and cost only 10% computation of baseline,and the equal error rate(EER) of the system in NIST LRE 2009 reduces 12.75%,15.89% and 7.33% relatively in 30s,10s and 3s respectively.更多还原