用于动态声纹密码认证系统的分类器和归一化的比较性研究

Comparison Study of Classifier and Normalization for Dynamic Voiceprint Password Authentication System

数据归一化和分类器种类对数据分类准确率有着重要影响,而归一化后的数据通过不同的分类器会产生不同的响应。在动态声纹密码认证系统中,对声纹与文本的得分分别采用向量规范化和线性变化法进行处理,利用概率神经网络和支持向量机分类器进行分类,同时讨论当不同归一化方法和不同分类器相结合时,其对系统识别率的影响。实验结果表明:支持向量机的分类器和线性变化归一化技术相结合时,支持向量机的非线性映射具有更为普遍和明显的优势,系统性能显著提升,识别率提升了11.69%。

Data normalization and the kinds of classifier have a great lntluence on me accuracy of classification, and the normalized data, however would produce different responses via different classifiers. Based on the vector normalization and linear variation normalization, the scores of voiceprint and test in dynamic voiceprint password authentication system are investigated, and the data also classfied with PNN （Probabilistic Neural Network） and SVM （Support Vector Machines）. In addition, the combined effects of different data normalization methods and different classifiers on the accuracy of classification are discussed. Experimental results indicate that the nonlinear mapping SVM has more universal and obvious advantages, and by combining SVM with Linear variation normalization, the recognition rate is significantly improved and increased by 11.69%.