Speech Emotion Recognition Based On Feature Selection

Speech emotion recognition is a hotspot problem in many areas, such as artificial intelligence, machine learning, etc. By finding speech characteristics that correspond to emotions, researchers use HMM, SVM and KNN classifier for speech emotion recognition. However, there may exist some problems in current studies, i.e. there are too many speech features, which affect the recognition rate and increase the computational time. In order to tackle these problems, this paper presents a feature selection method to reduce the dimension of features, preventing the "dimension disaster" and improving the recognition rate. This dissertation conducts the following research topics:(1) MMHC algorithm was applied for feature selection. Based on MMHC algorithm, a heuristic MMHC (h-MMHC) algorithm was proposed, which adopts a heuristic strategy that gradually adds features. In contrast to MMHC algorithm that handle all features at one time, h-MMHC can reduce the computational time.(2) h-MMHC was applied in speech emotion recognition problem and a speech emotion recognition framework was constructed based on h-MMHC algorithm. When using the framework to select features among MFCC dataset, the feature dimension was reduced to 21 characteristics from 180 characteristics, besides the classification recognition rate increase by 1% compared to non-feature selection methods.(3) Considering emotions in real world are not exclusive, we conduct feature selection for each emotion. The experimental results show that different emotions share a few characteristics. This results provide a theoretical support for the construction of emotional voice library.Through the above research, this dissertation proposed a heuristic MMHC algorithm for feature selection with the result of reducing computational time. When applying this algorithm to feature selection for speech emotion recognition problem, the recognition rate increased by one percent and shared characteristics of different emotions were found, which verified the non-exclusive characteristic of different emotions.