基于鉴别性向量空间模型的语种识别

Discriminative vector space model based language recognition

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(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有针对性,从而用较少的训练数据能取得较好的分类效果。在美国国家标准技术局(NIST)2009年语种识别测试中,D-VSMs只用了原训练数据的25%,计算量是传统并行音素识别器后接向量空间模型(PPRVSM)的10%,等错误率在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%。

Conventional language recognition tasks are limited by the need for large training datasets,in which most of the discriminative information is overlapped.Moreover,the non-language variabilities（such as channel and speaker differences） also affect the performance of language recognition systems.This paper describes a method using discriminative vector space models（D-VSMs） where the overlapping training information is automatically eliminated.Thus,every VSM is trained for one special situation,and the whole system has good performance.D-VSMs only use 30% of the training data of the baseline system and cost only 10% computation of the baseline with the equal error rate（EER） for the system in the National Institute of Standards and Technology（NIST） Language Recognition Evaluation（LRE） 2009 Database reduced 12.75%,15.89% and 7.33% in 30 s,10 s and 3 s tests.