Research On Speech Emotion Recognition Based On Fuzzy Clustering

Speech emotion recognition is an important part of Human-Machine Interaction,and is a new branch of Human Intelligence, which has a promising application in manydifferent fields. In this thesis, it reasearches Gaussian Kernel Fuzzy VectorQuantization (GKFVQ) for speech emotion recognition methodology, it contains thefollowing:1. In this thesis, it researches variance based Gaussian Kernel Fuzzy VectorQuantization algorithm (VGKFVQ) to recognize different unknown emotions. In thismethod, it uses the sample variance to replace the kernel width parameter in GaussianKernel function so that this parameter can be adjusted automatically and takes use ofthe distribution information of the sample. As well as it saves much time which wasspent to find the proper parameter out. Results shows this method has a goodrecognition rate than that of GKFVQ、LBG and FVQ.2. There is a local optimum issue which has maken recognition rate declined inVGKFVQ. Then in order to fix the local optimum problem, it researches joiningmodified Particle Swarm Optimum (PSO) in VGKFVQ (PSO-VGKFVQ) to classifyemotions. It uses modified PSO to look for the best particle at first and the best particleis regarded as the first generation clustering centers, the next step it still usesVGKFVQ method to train the code book. The experimental results indicate thisPSO-VGKFVQ algorithm can improve the performance for speech emotionrecognition compared with VGKFVQ algorithm.3. As for the problem happiness and angry are highly misrecognised between eachother when VGKFVQ algorithm gets into local optimum, it researches combiningSupport Vector Machine (SVM) with VGKFVQ method (SVM-VGKFVQ) repeat torecognize again. By fully using the advantage of SVM algorithm it uses SVM torecognize the results from VGKFVQ method in happiness and angry again to make acompensation for the local optimum deficiency in VGKFVQ. And the results showthat SVM-VGKFVQ method recognition rate is obviously higher than that of VGKFVQ.4. By using the result that position and the distribution of vectors in code books canimpact the discrimination of the speeches, moreover, it analyses the visualization of thefirst and the second dimension from happiness code books which have differentrecognition rate, it’s apparently that there are some vectors from bad recognition ratecode books which are far away from the others. So in this thesis, it researches usingsample mean and vector weighted to modify VGKFVQ algorithm to classifyemoitioms. It uses sample mean to substitute for every sample and vector weightedwhich can measure the discrimination ability of each vector in every code bookconstruct the modification item to discriminate unknown emotions. Compared withVGKFVQ algorithm, this method can slightly improve the recognition rate and have agood anti-noise ability.