基于SVM一对多得分规整的语种识别方法

A Normalization Method for Language Recognition Based on SVM one Versus All Scores

提出一种支持向量机(SVM)一对多得分规整的语种识别方法。通过对SVM得分进行规整,提高了各语种得分间的区分性,同时对分类效果较差的SVM分类器更鲁棒。仿真实验基于音素层特征的并行音素识别器后接向量空间模型(PPRVSM)的语种识别系统上进行,在美国国家标准技术署(NIST)2011年语种识别评测(LRE)30s数据集上的实验表明,提出的规整方法在语种识别性能评价指标EER和min DCF上相对提升2.6%-10.9%。

This paper presents a normalization method for language recognition based on Support Vector Machine（ SVM） one versus all scores. It improves the discrimination of scores between languages,and in the meanwhile,it is more robust for SVM classifiers whose classification results are not well. Experiments are carried out on Parallel Phoneme Recognizer followed by Vector Space Model（ PPRVSM） based on phonotactic features. Results on 2011 National Institute of Standards Technology（ NIST） Language Recognition Evaluation（ LRE） 30 s data set show that the proposed method achieves relative improvement of 2. 6%- 10. 9% in EER and min DCF.