Abstract—This paper presents an approach for robust automatic recognition of a speaker’s emotional states using a combination of prosody features (i.e. Pitch period, Short time energy and Short time zero crossings), quality feature (i.e. Formant frequencies) and the derived feature (i.e. MFCC -Mel-Frequency Cepstral Coefficients). Multilevel Support Vector Machine classifier is used for identification of four discrete emotional states namely boredom, anger, happy and sad for English language in Indian accent. The database was created by recording speeches of ten selected individuals who spoke sentences in all the four emotions. The experiment was carried out using both one versus all and one versus one methods of classification. The experiment was repeated using various kernel functions and by varying the feature vector length. The experiment aims at finding the kernel (such as linear, Gaussian radial basis function, multi-layer perceptron, polynomial and quadratic functions) and the feature vector length to get optimum results for the classification of each emotion.
 

Index Terms—prosody features, pitch period, zero crossings, energy, formant frequencies, Mel frequency cepstral coefficients, multi class support vector machine, statistical parameters, kernel functions, feature vector length

Cite: Keerthi R Shastry and A Sreenivasa Murthy, "Speaker Emotion Recognition Using Multiclass SVM for Evaluating the Best Kernel Functions and Feature Vector Length to Obtain Optimum Results," International Journal of Electronics and Electrical Engineering, Vol. 6, No. 4, pp. 76-80, December 2018. doi: 10.18178/ijeee.6.4.76-80