面向多语种语音数据的自适应情感识别算法研究

Speech emotion recognition is a very challenging research topic in affective computing. In many scenarios, we have to face the speech with unknown languages. So the research on language adaptation based speech emotion recognition is very valuable in theoretical aspects and practical applications. To emotional speech collected in different languages, although we might not understand their meanings, we can perceive the intrinsic emotional information. Based on our preliminary work in this area, in this project, we want to study the multi-language emotional speech, and deeply investigate the theories and algorithms of language adaptation based speech emotion recognition. The main research topics are given as follows. First, we will propose a low rank constrained transfer learning method to learn robust cross-language feature representations. Second, we plan to investigate the multi-view learning based cross-corpus emotion recognition via multi-view features. Finally, we will study two important parts of speech emotion recognition, i.e., classifier design and feature extraction, in which we will present a robust sparse regression emotion prediction method, while planning to extract new language-independent emotional features from the harmonic and noise model (HNM). This project clarifies the influencing factors of emotional speech in different languages from a new point of view, and will promote developments of speech emotion recognition techniques.

语音情感识别是情感计算中富有挑战性的研究课题，在许多应用场景中需要面对语种未知的情况，因而研究面向多语种语音数据的自适应情感识别算法具有重要的理论价值和现实意义。对于不同语种的语音，尽管我们可能不理解其语义，却可以感知其蕴含的情感信息。本项目旨在申请人原有语音情感识别工作的基础上，以多语种语音数据为研究对象，深入研究语种自适应的情感识别理论及方法。主要研究内容包括：首先，研究基于低秩约束的特征迁移学习方法来获取多语种鲁棒情感特征表示；其次，从特征多视角的角度出发研究基于多视角学习方法的跨语种情感识别；最后，对多语种情感识别的两个关键技术—分类方法和特征提取进行研究，提出基于鲁棒稀疏回归的情感预测方法；同时基于谐波噪声模型提取语种鲁棒能很好反映不同情感差异的特征。本项目从新的角度去阐明不同语种条件下影响情感语音的因素，其研究成果将为语音情感识别的发展产生积极的推动作用。

作为情感计算的一个重要分支，语音情感识别受到人们越来越多的关注。目前的语音情感识别主要针对某一特定语种数据库进行训练和识别，忽略了不同语种之间的差异。本项目通过挖掘不依赖于语种的情感特征，寻找适用于不同类型语音情感数据库的分类器，为鲁棒语音情感识别提供了一套切实有效的方案。具体包括：1）从情感特征迁移学习的角度出发，研究了基于子空间学习、稀疏编码、投影稀疏编码等一系列迁移学习相关方法，解决训练和测试语音情感特征分布不匹配引起的低识别率的问题；2）从情感特征提取的角度出发，研究了子空间学习和特征选择相融合的特征表示方法；3）从情感分类的角度出发，提出了基于稀疏判别回归的情感分类方法；4）从特征提取和分类相融合的角度出发，提出了基于回归和子空间迁移学习相结合的跨语种语音情感识别方法。本课题从新的角度去阐明不同语种环境下影响情感语音的因素，有助于进一步提高跨语种语音情感识别系统的性能。