论文部分内容阅读
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%。
Traditionally, the scale of training database in traditional language recognition is large, there is a great deal of overlap of discriminative information for language classification, and different sources of training data have different interference to training and testing. To solve these problems, a discriminative vector space model (D-VSMs) modeling method is proposed. D-VSMs can automatically filter the data of overlapping training information, so that each SVM training data is targeted, so that less training data can get better classification results. D-VSMs used only 25% of the original training data in the NIST 2009 Language Recognition Test, which calculated at 10% of the traditional parallel vector space model (PPRVSM) of the phoneme recognizer The error rate decreased by 12.75%, 15.89% and 7.33% respectively compared with the conventional PPRVSM under the test conditions of 30s, 10s and 3s.