作　　者： ; ; ;

机构地区： 清华大学信息科学技术学院计算机科学与技术系

出　　处： 《计算机研究与发展》 2003年第11期1561-1565,共5页

摘　　要： 支持向量机学习方法以结构风险最小化原则取代传统机器学习方法中的经验风险最小化原则,在有限样本的机器学习中显示出优异的性能.将这一新的统计学习方法应用到多媒体交互作用的研究中,用支持向量回归的方法由语音预测唇动参数.通过对语音的线性预测系数进行主分量分析,有效地压缩了声学特征参数的维数.结合交叉校验和最速下降优化方法,选择最佳的支持向量回归学习参数.在汉语0～9的任意数字串上对唇高参数的预测实验结果达到了均方误差0.0096,平均幅度误差7.2％及相关系数0.8的效果.这一结果优于一个文中优化过的人工神经网络所达到的性能,说明这一方法很有潜力. Unlike traditional machine learning which is based on empirical risk minimization principle, support vector machine (SVM) learning is based on structural risk minimization principle. SVM shows powerful ability in learning with limited samples. This new method is applied in the study of multimedia interaction and in predicting the mouth movement by speech based on support vector regression (SVR) . The audio parameters dimension is reduced by principle components analysis (PCA), and the optimal SVR learning parameters are selected based on cross-validation and steepest descent algorithm optimization. With the experiment on arbitrary Chinese digital numbers from 0 to 9, the prediction results reach 0.0096 in mean square error, 7.2% in absolute magnitude error, and 0.8 in linear correlation coefficient. It gives better results than that with optimized artificial neural network, which shows that the proposed method is quite promising.

关 键 词： 支持向量机 支持向量回归 线性预测系数 主分量分析 人工神经网络

领　　域： [自动化与计算机技术] [自动化与计算机技术]