基于模型自适应修正和协同决策的说话人鲁棒语音情感识别方法研究

项目名称： 基于模型自适应修正和协同决策的说话人鲁棒语音情感识别方法研究

项目编号： No.61003183

项目类型： 青年科学基金项目

立项/批准年度： 2011

项目学科： 无线电电子学、电信技术

项目作者： 毛启容

作者单位： 江苏大学

项目金额： 7万元

中文摘要： 语音情感识别是情感识别领域的关键研究问题之一。本项目着重研究了非个性化语音情感特征提取方法、基于模型参数自适应的非特定人语音情感识别方法以及多特征融合语音情感识别方法。主要成果包括：1) 提出了基于副语言的非个性化语音情感特征提取方法、基于多重分形的非个性化语音情感特征提取方法、基于声学上下文的语音情感特征提取方法，尽可能地屏蔽说话者个人说话特征对情感识别率的影响。2）提出了基于错分样本近邻支持向量优选的C-SVM在线自适应语音情感识别方法、基于特征分组的多核融合语音情感在线自适应识别算法，使得识别模型的参数能够适应待识别样本的变化。3）提出了基于可分度和支持度的模糊密度赋值融合识别算法、基于承诺和一致性系数的自适应模糊积分语音情感融合识别方法、基于增量流形学习的语音情感特征降维方法、基于过完备字典与PCA小样本语音情感识别方法，使得使用尽量少的、有效的语音情感特征，并融合每一类情感特征的优势，提高与人无关环境下的语音情感总体识别率。

中文关键词： 语音情感识别；说话者鲁棒；非个性化特征；模型参数自适应

英文摘要： Speech emotion recognition is one of key problems in the emotion recognition field. In this project, extraction methods of non-personalized speech emotion features, model adaption methods and fusion recognition methods of speech emotions have been researched. The main achievements include: 1) The non-personalized speech emotion feature extraction methods based on multi-fractal, paralanguage and speech context were presented. The features extracted by these approaches include much emotion information while avoiding the effect of different speakers. 2) Multiple kernel fusion on-line adaptive algorithm based on feature grouping, C-SVM on-line adaptive algorithm based on optimally selected nearest-neighbor support vectors around wrongly recognized samples were put forward. These algorithms were used into speech emotion recognition models, and the model parameters can be adapted with the change of the samples recognized. 3) Fusion recognition algorithm based on fuzzy densities determined with classification capability and supportability, speech emotion fusion recognition method with sample adaptive fuzzy density were proposed. These two fusion method can combine the merits of each type of speech emotion features. Speech emotion sparse representation method based on over-complete dictionary and PCA on the small sample corpus, dimensionality reduction method for speech emotion features based on incremental manifold learning were come up with. These approaches can get higher emotion recognition accuracy with speech emotion features as few as possible.

英文关键词： speech emotion recognition;speaker robust; non-personalized speech emotion features; model parameter adaption methods