基于OBSI和特征选择的语音情感识别算法

发布时间：2018-01-06 05:29

本文关键词：基于OBSI和特征选择的语音情感识别算法　出处：《大连海事大学》2017年硕士论文　论文类型：学位论文

【摘要】：语音情感识别利用语音信号中的情感信息识别人的不同情感状态,对解决因不良情绪造成的社会问题提供一种监测手段。本文首先寻找能够对语音情感有效识别的关键特征,再利用特征选择和区分不同情感的最优特征子集差异性,研究语音情感识别算法。本文主要工作如下:(1)给出一种适用于语音情感识别的改进频带划分的OBSI特征提取算法。将乐器识别领域的OBSI及其衍生特征OBSIR引入到语音情感识别中,并根据人耳听觉特性重新设计特征提取中间环节-倍频程滤波器组子频带频率范围。通过在EmoDB、CASIA和SAVEE数据库上的对比实验,验证了 OBSI特征和OBSIR特征用于语音情感识别的有效性,相比MFCC等常见特征有较好的识别能力;也验证了 OBSI和OBSIR特征与常见特征组合可以进一步提升语音情感识别效果。(2)给出一种基于样本特征值分布混淆度和MRMR结合的改进特征选择算法。根据数据集样本特征取值分布情况,首先定义并构建特征值混淆度统计量,以评估特征的重要程度,然后给出一种联合特征值混淆度和分类器的Wrapper式特征选择算法,并给出了相应的子集搜索策略,最后引入MRMR算法,对已选出的特征子集进行适当地维数约简。通过对比实验,验证了特征值混淆度统计量评估特征区分能力的有效性,也验证了本文给出特征选择算法能够选出较优的特征子集。(3)给出一种利用最优特征子集差异性的多级语音情感分类器的构建算法。根据情感区分度,使用若干个一对多的二分类SVM,首先对较为容易区分的情感进行区分,对于其中较难区分的情感使用随机森林进行识别。在多级语音情感分类器中,对用于训练子分类器的特征子集采取差异化策略,即子分类器用到的特征集合是从原始特征集合中选出针对特定情感类别的最优特征子集,根据先易后难的原则构建级联SVM和随机森林的改进多级语音情感分类器。实验验证了本文基于最优特征子集差异性来构建多级语音情感分类器算法的合理性。
[Abstract]:Emotion recognition of speech emotion recognition by human speech signal in different emotional states, provides a means of monitoring to solve the social problems caused by the bad mood. Firstly, to find the key features of effective speech emotion recognition, the feature selection and distinguish the different emotion optimal feature subset difference of speech emotion recognition algorithm. The main work is as follows: (1) OBSI characters are suitable for speech emotion recognition improved frequency division extraction algorithm. The musical instrument recognition OBSI and its derivative OBSIR features into speech emotion recognition, and re design according to the characteristics of human auditory feature extraction of intermediate links - octave filter group sub band frequency range. The EmoDB, CASIA experiments and SAVEE database on the OBSI and OBSIR features for speech emotion recognition is verified Effect, have better recognition ability than MFCC and other common characteristics; it also confirmed the OBSI and OBSIR features and common feature combination can further enhance the effect of speech emotion recognition. (2) an algorithm based on the eigenvalue distribution of sample selection algorithm to improve confusion characteristics and combine MRMR. According to the sample data sets feature value distribution. We first define and construct characteristic value confusion degree statistics to important degree evaluation characteristics, and then presents a joint eigenvalue Wrapper feature confusion degree and classifier selection algorithm, and gives the corresponding subset search strategy, finally introduced the MRMR algorithm, due to the dimensionality reduction of the feature subset is selected. Through the contrast experiment and verify the validity of the confusion degree statistics feature distinguishing ability of eigenvalues, which confirmed the given feature selection algorithm to select the optimal feature subset (3). The algorithm constructs a multilevel speech emotion classifier is presented by using the optimal feature subset difference. According to the emotion discrimination, with several of the two classification SVM, first of all is easy to distinguish the emotion to distinguish, for which it is difficult to distinguish the emotion using random forest recognition. In the multilevel speech emotion classifier and take the differentiation strategy of feature subset training sub classifier for feature sub classifier used is set to select the most feature specific emotion category subset from the original feature set, according to the principle of constructing frompointtoarea cascade SVM and random forest improved multilevel speech emotion classifier. Experiment verifies the rationality of the feature subset difference to construct multi classifier algorithm based on speech emotion.

【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.34

【相似文献】