应用于语种识别的加权音素对数似然比特征

发布时间：2018-08-30 15:57

【摘要】：语种识别的关键问题之一是提取语音信号中的语种鉴别性信息。近期,音素对数似然比(phone log-likelihood ratio,PLLR)的新特征被引入语种识别领域,并表现出了优异的性能。该文利用F比方法分析了PLLR特征向量各维的语种鉴别性大小,提出了加权音素对数似然比(weighted PLLR,WPLLR)特征,赋予PLLR特征中含有较多语种鉴别性信息的分量较高的权重。在美国国家标准技术署(National Institute of Standards and Technology,NIST)2007年语种识别测试集上的实验结果表明:相比于原PLLR特征,该文所提出的WPLLR特征在平均检测代价和等错率2个指标上都显著降低。
[Abstract]:One of the key problems of language recognition is to extract the discriminative information from speech signals. Recently, the new feature of phoneme logarithmic likelihood ratio (phone log-likelihood ratio,PLLR) has been introduced into the field of language recognition and has shown excellent performance. In this paper, we use the F ratio method to analyze the language discriminability of each dimension of the PLLR feature vector, and propose the weighted phoneme logarithmic likelihood ratio (weighted PLLR,WPLLR) feature, which gives a higher weight to the PLLR feature, which contains more language discriminant information. The experimental results on the 2007 language recognition test set of (National Institute of Standards and Technology,NIST show that compared with the original PLLR features, the proposed WPLLR features are significantly lower in both the average detection cost and the equal error rate.
【作者单位】：中国科学院声学研究所;国家计算机网络应急技术处理协调中心;
【基金】：国家自然科学基金资助项目(11461141004,91120001,61271426) 国家“八六三”高技术项目(2012AA012503) 中国科学院战略性先导科技专项(XDA06030100,XDA06030500) 中科院重点部署项目(KGZD-EW-103-2)
【分类号】：TN912.34

【相似文献】