【作者】 张乐；

【导师】 张雪英；

【作者基本信息】 太原理工大学 ， 信息与通信工程， 2017， 硕士

【摘要】 语音作为一种包含说话内容和情感状态的复杂信号,是人类进行交流和表达情感的有效形式。语音情感识别是计算机通过提取并分析情感语音的特征参数从而判别情感类别的一种信息处理技术,对于提高人机交互智能化程度具有重要意义。本文在语音情感识别的课题背景下,介绍了常用的语音库、情感特征和识别网络,将集合经验模态分解(Ensemble Empirical Mode Decomposition,EEMD)算法应用到语音情感特征提取中,提取了本征模态函数能量特征IMFE和边际谱幅值特征MSA,选择了IMFE、韵律特征、MFCC三种情感特征进行特征级融合,并提出了一种自适应融合核函数极限学习机(Extreme Learning Machine with Kernel,KELM)的决策级融合方法用于语音情感识别。本文所做的主要工作如下:(1)选择EEMD算法以非线性非平稳信号的处理方法提取情感语音特征。传统的情感特征提取方法均假定语音是短时平稳信号,针对传统方法的局限性,本文在EEMD算法分解语音信号的基础上提取了边际谱幅值特征MSA,并选择KELM为识别网络,基于柏林语音库设计仿真实验并对四种情感(高兴、悲伤、愤怒、中性)进行识别,通过与韵律特征、MFCC特征的识别结果对比,验证了MSA特征的有效性。(2)提出了一种基于EEMD算法的特征提取方法并应用于语音情感识别中。语音情感信号经EEMD算法分解为一组本征模态函数(IMF),通过Spearman Rank相关系数筛选出有效的IMF分量,并通过能量计算得到一个语音情感新特征IMFE,选择柏林语音库进行识别,并与韵律特征、MFCC特征的识别性能对比,结果表明IMFE可以有效识别情感,且对负性情感的识别效果最优。(3)将特征级数据融合应用于语音情感识别。针对单一语音情感特征识别效果不好的问题,本文选择了IMFE特征、韵律特征、MFCC特征进行融合,设计实验将这三种特征的不同组合分别输入到分类器中,在柏林语音库仿真并与输入的单一特征识别结果对比,结果表明特征融合在一定程度上提高了识别性能,证明了三种特征具有互补性,但也因为特征维数的简单相加造成了特征融合在部分情感的识别率低于单一特征识别率的问题。(4)提出了一种基于融合KELM的语音情感识别新方法。针对单一特征、单分类器识别性能不佳的问题,本文将决策级数据融合应用于语音情感识别的研究中,首先提取三种语音情感特征,并分别训练对应的单分类器,同时把单分类器的数值输出统一转化成概率输出;然后通过制定的决策策略得到测试集的自适应权值,决策策略依据概率矩阵而定;最后对各单分类器的输出概率线性加权并判别输出。选择柏林语音库进行识别,结果表明融合KELM在单一情感和整体的识别率均达到最优,优于单一特征、特征融合和常用决策策略的性能,是一种有效的语音情感识别方法。更多还原

【Abstract】 As a complex signal which contains the content and emotion,speech is an effective form of communication and expression.Speech emotion recognition is one kind of information processing technology that can judge the speech emotion state by extracting and analyzing the characteristic parameters of emotional speech,which is of great significance to improve the intelligence level of human computer interaction.Based on the background of speech emotion recognition,three parts of speech emotion recognition systems are introduced,namely,the common speech database,emotional features and identification network.The EEMD algorithm is applied to speech feature extraction,we extract the intrinsic mode function energy feature(IMFE)and marginal spectrum amplitude feature(MSA),and select the IMFE,prosodic features and MFCC for feature level fusion,and propose an adaptive fusion kernel extreme learning machine(KELM)of the decision level fusion method for speech emotion recognition.The research work what has done is as below:(1)The EEMD algorithm is adopted to extract the speech emotion feature by the nonlinear non-stationary signal processing method.The traditional speechfeature extraction methods assume that the signal is a short-time stationary signal.Aiming at the limitation of traditional methods,the marginal spectrum amplitude characteristics are extracted based on the EEMD decomposition of the speech emotion signal,and the KELM is used as recognition network.The experiments were conducted in EMO-DB to identify four emotions(joy,sadness,anger and neutral),and the validity of the MSA was verified compared with the results of prosodic features and MFCC.(2)A feature extraction method based on EEMD algorithm is put forward and applied to speech emotion recognition.Emotional speech signal is decomposed into a group of IMF by EEMD,the Spearman Rank correlation coefficient is used to screen the effective component of IMF,and a new feature of speech emotion named IMF energy(IMFE)is obtained by calculating the energy.The results of simulation in Berlin speech database and comparison with the recognition rate of prosodic features and MFCC show that IMFE can effectively identify emotion and the recognition performance of negative emotion is the best.(3)The feature level data fusion is applied to emotional speech recognition.In order to solve the problem of poor recognition performance of the single emotion feature,the IMFE,prosodic features and MFCC are chosen,and different combinations of these three kinds of features are used as the input of the classifier.The results of simulation in Berlin speech database and comparison with the results of single character show that the feature levelfusion improves the recognition performance in a certain extent and three features are complementary,but because of the simple addition of the feature dimension,the recognition rate of feature fusion is lower than that of the single feature in partial emotion.(4)A new method of speech emotion recognition on the basis of fusion KELM is proposed.In order to solve the problem of poor recognition performance of single feature and single classifier,the decision fusion is applied to speech emotion recognition.Firstly,three kinds of speech emotion features are extracted and the corresponding single classifier is trained respectively.At the same time,the numerical output of KELM is transformed into probability output.Then the adaptive weights of the test set are obtained by the decision strategy.The fusion strategy is made on the basis of probability matrix.Finally,the output value is obtained by linear weighting the output probability of each single classifier.The results of simulation in Berlin speech database show that the fusion KELM achieves the best recognition rate in one single emotion and the whole,it is superior to the single feature,feature fusion and the common decision strategy,so fusion KELM is an effective method for speech emotion recognition.更多还原