基于语谱图特征的语音情感识别研究

发布时间：2018-04-13 05:21

本文选题：语谱图 + Gabor滤波器　；参考：《太原理工大学》2017年硕士论文

【摘要】：目前,在语音情感识别领域,研究者经常选用的情感语音特征有音质特征、频域特征、时域特征等特征,而关注语音时域-频域相关性的研究则相对较少,发展较晚。本文利用语谱图反映语音时域-频域相关性的特点,提取语谱图多种不同的纹理特征,将所提特征应用于语音情感识别。通过实验分析,部分语谱图纹理特征取得了比较好识别率,验证了基于语谱图特征的语音情感识别方法的可行性。本文主要以语谱图为基础对语谱图纹理特征的提取和分类进行了研究,主要完成的工作包含以下几部分:(1)介绍了语音情感研究的背景、意义、发展现状以及常用的图像纹理特征提取方法和纹理特征分类方法。(2)利用Gabor小波分别结合灰度共生矩阵方法、Tamura方法、局部二值模式(LBP)方法提取语谱图特征。(3)提出了改进的局部二值模式方法、融合语谱图LBP特征和局部Hu矩特征方法,并以柏林语音库为基础进行情感语音识别,实验结果显示不同情感的综合识别率有了比较明显的提升。(4)实现了支持向量机和K近邻分类方法对所提特征进行分类识别,并研究了语谱图灰度共生矩阵特征、Tamura特征、LBP特征、改进的LBP特征、融合LBP特征与局部Hu距特征的情感识别效果,以及融合特征权值对识别率的影响。经过比较不同方法下的识别率,改进后的LBP方法与融合语谱图LBP特征与局部Hu距特征方法取得了较好的实验结果。
[Abstract]:At present, in the field of speech emotion recognition, researchers often choose affective speech features, such as phonological characteristics, frequency domain features, time domain features, etc. However, the research on the correlation between speech time domain and frequency domain is relatively few, and the development is relatively late.In this paper, the feature of correlation between speech time domain and frequency domain is reflected in the speech spectrum, and different texture features are extracted from the spectrum map, and the proposed features are applied to speech emotion recognition.Through experimental analysis, some of the texture features of spectral images have achieved a good recognition rate, which verifies the feasibility of the speech emotion recognition method based on the spectral image features.In this paper, the extraction and classification of texture features are studied on the basis of spectrogram. The main work includes the following parts: 1) introduce the background and significance of speech affective research.The current situation and the common methods of image texture feature extraction and texture feature classification. (2) using Gabor wavelet and gray level co-occurrence matrix method, respectively, Tamura method is used.This paper presents an improved local binary pattern method, which combines LBP features with local Hu moments features, and implements emotional speech recognition on the basis of Berlin phonetic corpus.The experimental results show that the comprehensive recognition rate of different emotions is significantly improved. (4) support vector machine and K-nearest neighbor classification method are used to classify and recognize the proposed features, and the gray level co-occurrence matrix feature Tamura feature and LBP feature of the spectral image are studied.The improved LBP feature, the emotion recognition effect of fusion LBP feature and local Hu distance feature, and the influence of fusion feature weight on recognition rate.By comparing the recognition rates of different methods, the improved LBP method, the fusion spectral LBP feature and the local Hu distance feature method have obtained good experimental results.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.34

【相似文献】