基于压缩感知的鲁棒性语音情感识别研究

Spoken emotion recognition is currently an active research topic in the fields such as signal processing, pattern recognition, artificial intelligence, human machines interaction, etc. The emotional speech signals in the natural scenery are usually disturbed by the noise and thus contain different levels of noise. Therefore, it is more practical and meaningful for the researches on robust spoken emotion recognition under the noise background. Based on the compressive sensing theory, this project launches the researches on robust spoken emotion recognition under the noise background. The research topic in this project includes three main parts: (1)the sparse representation and recovery of emotional speech signals in the noise background;(2)the sparse discriminant manifold learning methods based on the compressive sensing theory; (3)the design of sparse classifier for robust spoken emotion recognition. The novelty of this project lies in the integration of the spoken emotion recognition theory, the compressive sensing theory, the pattern recognition theory, and so on. This project originally proposes a new method of robust spoken emotion recognition based on the compressive sensing theory, in an effort to solve the problem that how to effectively improve the performance of spoken emotion recognition in the noise background. The researches on this project will enrich the pattern recognition theory,the machine learning theory, as well as the affective computing theory,and present a new approach of robust spoken emotion recognition. The research idea and means in this project have important scientific values to drive and accelerate the development of some disciplines such as human machines interaction, artificial intelligence, affective computing, and so on.

语音情感识别是当前信号处理、模式识别、人工智能、人机交互等领域的研究热点。自然环境中的情感语音信号通常都会受到噪声的干扰，包含了不同程度的噪声。因此，对于噪声背景下的鲁棒性语音情感识别方面的研究更接近实际，更具有研究意义。本项目基于压缩感知理论开展噪声背景下的鲁棒性语音情感识别研究，主要包括：（1）噪声背景下的情感语音信号的稀疏表示及重构；（2）基于压缩感知理论的稀疏判别流形学习方法；（3）面向鲁棒性语音情感识别的稀疏分类器的设计。项目的创新之处在于整合语音情感识别理论、压缩感知理论、模式识别等理论，首次提出基于压缩感知的鲁棒性语音情感识别方法，以解决噪声背景下如何有效提高语音情感识别性能的问题。本课题的研究将丰富模式识别、机器学习、情感计算等理论，为鲁棒性语音情感识别的研究提供新的解决途径，其研究思想和方法对推动和促进人机交互、人工智能、情感计算等学科的发展具有重要的科学意义。

本项目从压缩感知理论入手，对情感计算领域的一个重要内容——语音情感识别在稀疏表示、特征降维、分类器设计等方面开展相关研究。在稀疏表示方面，对基于字典学习的稀疏表示方法进行了探索；在特征降维方面，将压缩感知理论与流形学习方法相结合，提出了局部稀疏判别投影等方法；在分类器设计方面，提出了局部敏感的核稀疏表示分类器等方法。项目组成员三年来共发表论文15篇，包括期刊论文14篇（其中SCI期刊6篇，EI期刊4篇），会议论文1篇，获得授权发明专利1项。