Study Of Speech Emotion States Fuzzy Recognition

Speech emotion recognition technology is an emerging branch of artificial intelligence, which has a wide range of applications in building a harmonious human-computer interaction environment and safety oversight system automatically. This technology has a great significant for human progress. Speech emotion recognition progress includes voice signal preprocessing, speech emotion feature extraction and emotion classification. It is difficult to identify the emotion state exactly, because the semantic variable of emotion has fuzzy and uncertainty information. It has a low recognition rate for the high dimension of emotional features. In this thesis, it researches a new method for speech emotion recognition based on fuzzy theory. The main contents are as follows:1. In this thesis, we research the speech signals preprocessing steps including endpoint detection, pre-emphasis, framing and windowing, extraction emotion features and used KPCA(Kernel Principal Component Analysis) to reduce the feature dimensions. We improve recognition performance by reducing the feature redundancy and the computational learning algorithms. Experimental results show that the method based on kernel principal component analysis feature selection makes recognition results improved.2. In this thesis, we research the speech emotion recognition method based on fuzzy support vector machine algorithm(FSVM). It improves support vector machine learning by combining the fuzzy theory. It solves small sample, nonlinearity, high dimensionality, local minimum by its generalization ability. It improves speech emotion recognition rate by weakening the influence of noise and isolated points in the fuzzy classification method.3. In this thesis, we research the speech emotion recognition method based on adaptive fuzzy C-means algorithm(AFCM). This algorithm can detect a data set in different geometries clustering. This method for different data clustering using different radius, make up the FCM uses less the same radius, improved speech emotion recognition rate. When using Mahalanobis distance function and fuzzy weighted index m = 2, We found that it has a high recognition for the four emotion state. We found that AFCM algorithm has better recognition performance than fuzzy C-means algorithm(FCM).4. In this thesis, we research the speech emotion recognition method based on fuzzy kernel vector quantization algorithm(FKVQ) and fuzzy kernel entropy vector quantization(FKEVQ).in FKEVQ algorithm, it quantizes the similar data into code book by introduced in the code word vector quantization. It has a better distinction by using kernel mapping the input space into a high dimensional feature space. It has a high data distinguished ability by introduced fuzzy entropy, which can balance membership function. When the number of code word C=22, fuzzy weighted index m=1.1 and the width of the Gaussian kernel δ=5, we found that the FKVQ algorithm performance is optimal. In the FKEVQ algorithm, the error rate becomes high by increase the initial fuzzy entropy value. When initial fuzzy entropy value λ=0.06, FKEVQ algorithm performance is optimal.