Deep Affective Interaction Model In Emotion Dimension

Speech emotion recognition mainly improves the communication between human and machine by effectively capturing and processing speech emotional information,ensures the natural interaction between human and machine,and makes the machine more “compassionate” in human-computer interaction.Emotion recognition model can realize the correct classification of emotion,as an effective means of speech emotion recognition.In view of the fact that the existing emotion model only divide the emotional state from the space and the interaction between emotions is neglected.Therefore,this paper proposes a deep affective association model which combines multi-layer Restricted boltzmann machine with emotion interactive cognitive network.In order to study the different distance relationship between emotions,the distance similarity algorithm is introduced into the deep affective association model.The calculated values of Manhattan distance,Chebyshev distance and Cosine similarity are used as the correlation degree between emotions,and the effectiveness of the deep affective association model proposed in this paper is verified by experiments.The main contents of this paper are as follows:(1)The speech emotion recognition system is introduced.Aiming at the division of emotion,two emotion description models are introduced in detail:discrete emotion description model and dimensional emotion description model;Secondly,several commonly used emotion databases are introduced,according to the research content of this paper,TYUT1.0 and CASIA emotion database are selected as the source of experimental data;Emotional speech features are extracted: traditional emotional acoustic features(prosodic features and MFCC)and nonlinear features(nonlinear attributes and nonlinear geometry).Finally,several commonly used emotion recognition networks are introduced: Hidden Markov model,Gaussian mixture model,Support vector machine and Deep belief network.(2)A deep affective association model is established and the validity is verified.This paper introduces the PAD three-dimensional emotional space model,explains the relationship between emotions from the perspective of emotional dimension,and takes the distance between emotions as the weightbetween emotional categories to establish an associated cognitive network.Introduces the multi-layer restricted Boltzmann machine,through the multi-layer restricted Boltzmann machine to solve the problems of many dimensions of emotional features and large redundancy,and takes the output weight of the multi-layer restricted Boltzmann machine as the weight between the input and output of the associated cognitive network to construct a deep affective association model.Through the experimental comparison between the deep belief network and the deep emotional association model,the experimental results show that the average recognition rate of deep emotional association model is 6.06% higher than that of Deep Belief Network.It has a better emotional recognition performance.The results prove that the deep emotion association model has strong superiority and universality in speech emotion recognition,and can well reflect the interaction between emotions.(3)The deep model based on distance similarity is studied Combined with the similarity algorithm.This paper introduces three algorithms based on distance similarity: Manhattan distance,Chebyshev distance and cosine similarity,which represent the distance between emotions in PAD three-dimensional emotional space,and take it as the correlation degree between emotions in relational cognitive network.Comparative experiments with deep belief networks are carried out to verify the effectiveness of the three algorithms in the construction of deep affective association model.Finally,the three algorithms are compared with Euclidean distance to study the advantages of distance similarity algorithm in speech emotion recognition.The results show that Chebyshev distance shows better recognition performance on TYUT1.0database,and Cosine similarity recognition results are better on CASAI database.