论文部分内容阅读
针对语种识别中大规模数据库的训练问题,提出一种基于局部多样性建模的向量空间模型。首先将训练数据库分成若干个小数据库,然后利用每个小数据库来训练不同的向量空间模型,最后对不同的模型进行加权组合。为了有效地对不同模型进行组合,需要对模型的加权系数进行优化。对模型组合算法从理论上进行推导,在模型权重与分数线性融合系数之间建立起对应的数学关系,并提出采用逻辑回归方法对不同模型的权重进行估计。在美国国家标准技术局(NIST)2009年度语种识别测试库上的实验结果表明:所提方法不仅能够处理大规模的训练数据,而且相比传统方法识别性能也有了一定程度的提高,系统的等错误率在30 s、10 s和3 s的测试条件下分别下降了8.44%、5.91%以及3.45%。
Aiming at the training of large-scale databases in language recognition, a vector space model based on local diversity modeling is proposed. First, the training database is divided into several small databases, and then each small database is used to train different vector space models. Finally, different models are weighted and combined. In order to effectively combine different models, the weighting coefficients of the model need to be optimized. The model combination algorithm is theoretically deduced, and the corresponding mathematical relationship is established between the model weight and the fractional linear fusion coefficient, and the logistic regression method is proposed to estimate the weight of different models. The experimental results on the National Institute of Standards and Technology (NIST) 2009 Language Recognition Test Database show that the proposed method can not only process large-scale training data, but also improve the recognition performance compared with the traditional methods. The error rate decreased by 8.44%, 5.91% and 3.45% respectively under the test conditions of 30 s, 10 s and 3 s.