Emotion Recognition Algorithm Research Based On The Variety Of Features Of Speech Signal

Affective computing is the calculation that associated with emotion, from emotional or emotional effect. Its purpose is to give intelligent machine perception, understanding and expression ability of various kinds of emotional state. Emotion recognition is the key of emotion computing research content. It is the important foundation to realize the harmonious human-computer interaction. In the field of emotion computing, the physiological signals, facial expression, posture signal. The speech signal has got the attention of the researchers because that it is the main carrier of emotional information expression, and it is convenient to obtain. Therefore, extracting the effective parameters that can express the emotion and seeking effective model for the classification of speech signal is still the hot field that scholars constantly chasing.Based on the above conditions, this topic aimed at seeking the emotional characteristic parameters and the emotional model for four kinds of emotions (happy, angry, sad, fear) recognition. First of all, sample entropy has better distinguish in speech emotion recognition, this paper extract the sample entropy with traditional acoustic parameters as the effective parameters for emotion recognition; Then, in order to improve the system identification performance, this paper presents a neural network speech emotion model based on the traditional BP network, combined with PCA algorithm based on contribution analysis. The analysis of the experimental results show that the new fusion parameters can improve the performance of the classifier compared with the common characteristic parameters, and the emotional model this research adopted improve the recognition efficiency. The concrete research content is as follows:(1) The signal acquisition. According to the mechanism of speech signal、 emotion classification methods in the world and the typical domestic and foreign emotional speech database, and ultimately determine this emotion classification method, and the speech database of this paper.(2) Early treatment. After pretreatment, in view of the speech signal endpoint effect, this study proposes a two-stage discriminant algorithm based on sample entropy is:It introduces the concept of sample entropy, on the basis of the two-stage discriminant method, and uses the FCMC algorithm and BIC algorithm to determine the double threshold, as the first stage, to judge the load-point of signals preliminary; At the same time, chooses the correction of the zero rate as a secondary criterion, to determine the load-point of signals ultimately.(3) The feature extraction and selection. This paper extracts the sample entropy, traditional acoustic parameters (speed, energy characteristics, pitch frequency, MFCC) and their statistical parameters as the effective parameters for emotion recognition. This paper uses PCA algorithm to implement the original feature dimension reduction. And it can be used to get the most contracted vector set, to reduce the complexity of the network model, and to reduce the training time.(4) Emotional recognition. This paper presents a neural network speech emotion model based on the traditional BP network, combined with PCA algorithm based on contribution analysis. Based on the fusion feature, the emotional model was carried out on the 800 samples of four kinds of emotional state identification. Through analysis of experimental results, this paper suggests that the new fusion parameters have good recognition effect compared with the common characteristic parameters, and the emotional model this research adopted have a better effects.Considering the above analysis, the fusion feature that contains sample entropy feature and traditional acoustic parameters, combined with the emotion recognition based on PCA neural network model, can establish an effective speech emotion recognition system.