摘要
针对稀疏表示识别方法需要大量样本训练过完备字典且特征冗余度较高的问题,提出了结合过完备字典学习与PCA降维的小样本语音情感识别算法.该方法首先用PCA降维方法将特征降维,再将处理后的特征用于过完备字典训练与稀疏表示识别方法,从而给出了语音情感特征的稀疏表示方法,并确定了新算法的具体步骤.为验证其有效性,在同等特征维数下,将方法与BP、SVM进行比较,并对比、分析语音情感特征稀疏化前后对语音情感识别率、时间效率以及空间效率的影响.试验结果表明,所提出方法的识别率比SVM与BP高;与采用稀疏化前的特征相比,稀疏化后的特征向量更便于处理,平均识别率提高约15%,时间效率提高近原来的1/2,空间效率提升近原来的1/3.
To solve the problems of sparse representation recognition algorithm with a large number of samples to train over-complete dictionary and high degree of feature redundancy, the recognition algorithm of speech emotion on small samples by over-complete dictionary learning and PCA dimension reduction was proposed. The feature dimensions were reduced by PCA method, and the over-complete dictionary was trained to identify samples. The speech emotional feature sparse representation was given to determine the steps of the proposed algorithm. The proposed algorithm was compared with BP and SVM for the same feature dimensions. The effects of the feature sparse representing on speech emotion recognition rate, time efficiency and space efficiency were analyzed. The experimental results show that the recognition rate of the proposed algorithm is better than those of SVM and BP, and the sparse feature is easy to handle. Using sparse features, the average recognition rate is increased by nearly 15%, and the time efficiency is improved by nearly 50% with increased space efficiency of 33%.
出处
《江苏大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2013年第1期60-65,共6页
Journal of Jiangsu University：Natural Science Edition
基金
国家自然科学基金资助项目(61003183)
江苏省自然科学基金资助项目(BK2011521)
江苏大学高级人才基金资助项目(10JDG065)
关键词
语音情感识别
过完备字典
PCA降维
稀疏表示
识别率
recognition of speech emotion
over-complete dictionary
PCA dimension reduction
sparserepresentation
recognition rate
