Research On Speech Emotion Recognition Based On GA Feature Fusion And Decision Tree

With the rising demand for human-computer interaction,speech emotion recognition technology attracts many scholars to conduct research.At present,researchers mainly focus on speech signal processing,emotion feature set extraction,emotion feature selection and fusion,classifier construction.Considering that emotion feature sets and classifiers play a key role in the final result of speech emotion recognition,this thesis focuses on these two points: How to fuse different features to achieve high-quality emotion recognition and how to build an effective classification structure to obtain adaptive feature sets for current emotion categories.The main research work of this thesis are as follows:(1)This thesis investigates the current situation and future development of speech emotion recognition technology,and then analyzes each module of current mainstream speech emotion recognition technology.A speech emotion recognition system includes database,emotion feature,feature selection,feature fusion method,classifier and other modules.Further,this thesis completes the performance evaluation of each module through experimental simulation.On the basis of the above,this thesis analyzes the problems in speech emotion recognition,finds out the potential solutions and provides theoretical for the follow-up research work.(2)In order to solve the problem that single category featurs can not fully represent the emotional information,this thesis proposes a speech emotion recognition method using genetic algorithm to combine deep bottleneck features and acoustic features.On the one hand,this method uses MFCC,pitch frequency,energy,zero crossing rate to represent the acoustic change information of different emotions in the speech.On the other hand,this method designed a DNN to extract the deep bottleneck feature of speech,which is used to make up for the lack of information associated with classification labels in acoustic emotional features.Next,GA is introduced to search the contribution weight of the two types of features.The search results are used to realize the combination of the two types of features.Finally,SVM is used to realize the training and classification.The experimental results show that the fusion method based on genetic algorithm can obtain more distinguishing feature set,which has higher recognition performance than single feature set.(3)Since the most adaptive feature sets related to emotions are different,this thesis uses more adaptive feature sets for current emotions to improve the recognition results.Based on the above emotion recognition strategy,the tree and direct structure of speech emotion recognition methods are proposed.Speech emotion recognition based on tree structure uses the same optimization target for emotions with similar characteristics.The obatained feature sets are more suitable for these emotion categories.Speech emotion recognition based on direct structure uses different adaptive feature sets for each type of emotion.Finally,high-quality classification of each type of emotion can be achieved.The experimental results show that the two structures of speech emotion recognition system can improve the performance.Tree structure is better than direct type structure in time complexity,while direct type structure can achieve higher performance.