摘要
为提高C-SVM的泛化性能,提出一种基于特征分组的多核融合在线自适应识别算法.此算法首先把特征按照待识别样本集的特性分为若干组,然后根据各组特征的特性采用不同的核函数训练C-SVM模型,并分别把各个模型支持向量间的相似度作为其权重系数,通过自适应样本不断调整权重系数和模型参数,使得C-SVM模型的参数能够随着待识别样本特性的变化而自适应地变化.将此算法应用于非特定人语音情感识别系统,与RBF核、多项式核和Sigmoid核的对比证明了多核融合在线自适应识别算法的优越性,通过与中性语句归一化方法相比证明了本文算法的有效性和稳定性.
In order to improve the generalization performance of C-SVM,a novel multiple kernel fusion on-line adaptive algorithm based on feature grouping is proposed.In this algorithm,the features of samples are grouped according to their attributes to form respective train-set,then different kernel functions are chosen to match different groups and train C-SVM models respectively.After models trained,the similarities of support vectors in models are computed as the weight coefficient of each classes,and the weight coefficient and parameters of models are adjusted by the adaptive samples.The parameters of C-SVM are gradually changed along with the change of the features of recognized samples.Finally,the optimal C-SVM model is gotten.This algorithm is applied into the independent-speaker speech emotion recognition system,the optimality is revealed by the comparison with RBF,multinomial function and Sigmoid function.The validity is proofed by the comparison with the method of normalizing neutral samples.
出处
《小型微型计算机系统》
CSCD
北大核心
2013年第3期585-589,共5页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61003183)资助
江苏省自然科学基金项目(BK2011521)资助
江苏省高校自然科学研究面上项目(09KJB520002)资助
江苏大学高级人才基金项目(10JDG065)资助