摘要: 非线性流形学习可以准确反映现实非线性数据本质并进行较好的降维,但在语音情感识别过程中难以有效处理不断增加的语音数据集,也不能充分利用训练过程中的情感特征信息。针对上述情况,提出一种基于增量流形学习的语音情感特征降维方法。该方法利用等距映射将训练样本特征维数降至目标维数后,通过增量流形学习的方法分批求得测试样本的低维特征。实验结果表明,相比同类方法,该方法具有较低的运算复杂度和较高的识别率。
关键词:
语音情感识别,
增量流形学习,
特征降维,
等距映射,
支持向量机
Abstract: Nonlinear manifold learning has the advantages of accurately reflecting the real nature of nonlinear data and effective dimension reduction. But it can not effectively handle ever-increasing speech data set and can not make full use of information got from the training process in speech emotion recognition. This paper presents a speech emotional feature dimension reduction method based on incremental manifold learning, which employs Isometric Mapping(ISOMAP) to reduce the feature dimension of training sample to target dimension, and obtains the low-dimensional features of test sample by using the incremental manifold learning method. Experimental result shows that the method has lower computational complexity and achieves better recognition performance.
Key words:
speech emotion recognition,
incremental manifold learning,
feature dimensionality reduction,
Isometric Mapping(ISOMAP),
Support Vector Machine(SVM)
中图分类号:
王海鹤, 陆捷荣, 詹永照, 毛启容. 基于增量流形学习的语音情感特征降维方法[J]. 计算机工程, 2011, 37(12): 144-146.
WANG Hai-He, LIU Cha-Rong, DAN Yong-Zhao, MAO Qi-Rong. Dimensionality Reduction Method for Speech Emotional Feature Based on Incremental Manifold Learning[J]. Computer Engineering, 2011, 37(12): 144-146.