摘要
针对支持向量机不能直接处理动态时间序列的语音数据问题,提出一种基于Fisher分值法的特征提取方法。Fisher分值法可以有效地进行特征向量的定长转换,使得支持向量机可以在整体语音序列上进行分类,从而提高系统的识别率。仿真实验结果表明,该方法在不影响系统识别速度的情况下,具有较高的识别性能。
Aiming at the restriction of SVM in working with fixed-length vectors, a novel feature extraction approach based on Fisher score was proposed. Fisher score can convert the input vectors of variable length into fixedlength vectors effectively. By doing so, SVM could classify on whole sequence. Experiment results showed that this approach achieved good performance and had no impact on the speed of system.
出处
《科学技术与工程》
2008年第21期5854-5857,共4页
Science Technology and Engineering
关键词
语音确认
特征提取
Fisher分值
支持向量机
高斯混合模型
speech sounds verification feature extraction Fisher score support vector machine Gaussian mixture model