选择性AdaBoost SVM语音情感识别算法的研究

发布时间：2018-07-21 11:21

【摘要】：作为人机交互技术中重要的成员之一，语音情感识别技术被广泛应用在教育、医疗、通信、计算机、自动化等行业。同时，语音情感识别涉及的知识面很广，涵盖了计算机科学与技术、模式识别、语音学、心理学、统计学和信号处理等学科。具有良好的研究基础和广阔的发展前景。目前，语音情感识别的相关研究已经取得很多的成果，同时也存在着许多研究困难。通过改进分类算法准确率可以提高语音情感识别产品和系统的性能，使其提供更好的服务和用户体验，提高一些行业的工作质量和效率，对促进行业发展具有重要意义。在语音情感识别方面，SVM算法具有很好的分类性能，而AdaBoost算法可以进一步提升SVM算法的分类准确率。本文以SVM和AdaBoost算法为基础，提出一种新的集成学习算法，即选择性AdaBoostSVM算法，算法思路为：首先使用AdaBoost算法训练若干个SVM分类器，再通过Kmeans算法对这些分类器进行聚类，得到若干个代表分类器，，然后对每一个测试样本，均使用Knn算法从训练集中找出其最近邻的若干训练样本，并将这些训练样本放入代表分类器中测试，最后选出测试准确率最高的分类器作为当前测试样本的最终分类器。本文在EMO-DB德语语音库、CASIA中文语音库和SAVEE英语语音库下测试算法效果。先找出三个语音库在五倍交叉验证和十倍交叉验证下的最优SVM参数，再分别对三个语音库作五倍交叉验证和十倍交叉验证的测试。实验结果表明，本文算法对三个语音库的分类准确率均有提升。在五倍交叉验证中，本文算法对三个语音库的分类准确率较单一SVM算法分别提升了1.86%、1.51%和3.77%，较AdaBoostSVM算法提升了0.35%、0.78%和0.13%，分类准确率达到87.56%、81.75%和76.75%。在十倍交叉验证中，本文算法对三个语音库的分类准确率较单一SVM算法分别提升了1.46%、0.74%和1.86%，较AdaBoostSVM算法提升了0.21%、0.37%和1.86%，分类准确率达到87.29%、81.20%和76.51%。说明本文提出的选择性AdaBoostSVM算法在提升语音情感识别准确率上是可行的。
[Abstract]:As an important member of human-computer interaction technology, speech emotion recognition technology is widely used in education, medical treatment, communication, computer, automation and other industries. At the same time, speech emotion recognition involves a wide range of knowledge, covering computer science and technology, pattern recognition, phonetics, psychology, statistics and signal processing and other disciplines. It has a good research foundation and broad development prospect. At present, the research of speech emotion recognition has made a lot of achievements, but also there are many difficulties. By improving the accuracy of classification algorithm, we can improve the performance of speech emotion recognition products and systems, make them provide better service and user experience, and improve the working quality and efficiency of some industries, which is of great significance to promote the development of the industry. SVM algorithm has good classification performance in speech emotion recognition and AdaBoost algorithm can further improve the classification accuracy of SVM algorithm. Based on SVM and AdaBoost algorithm, a new ensemble learning algorithm, selective AdaBoost SVM algorithm, is proposed in this paper. The idea of the algorithm is as follows: firstly, several SVM classifiers are trained by AdaBoost algorithm, and then these classifiers are clustered by Kmeans algorithm. A number of representative classifiers are obtained, and then, for each test sample, the Knn algorithm is used to find some training samples from the training set, and the training samples are put into the representative classifier for testing. Finally, the classifier with the highest test accuracy is selected as the final classifier of the current test sample. In this paper, the algorithm effect is tested in EMO-DB German language corpus, CASIA Chinese language corpus and SAVEE English phonetic database. The optimal SVM parameters of the three speech banks under five times cross validation and ten times cross validation are found out first, and then the five times cross validation and ten times cross verification tests of the three speech banks are performed respectively. The experimental results show that the classification accuracy of the algorithm is improved. In the five-fold cross-validation, the classification accuracy of the three corpora in this paper is 1.86% and 3.77% higher than that of the single SVM algorithm, 0.35% and 0.13% higher than that of the AdaBoost SVM algorithm, respectively, and the classification accuracy is 87.56% and 76.75%. In the 10-fold cross-validation, the classification accuracy of the three corpora in this paper is 1.46% and 1.86% higher than that of the single SVM algorithm, 0.21% and 1.86% higher than that of the AdaBoost SVM algorithm, and the classification accuracy is 87.29% 81.20% and 76.51% higher than that of the AdaBoost SVM algorithm. It shows that the proposed selective boost SVM algorithm is feasible in improving the accuracy of speech emotion recognition.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.3

【参考文献】