Mandarin emotion recognition based on multifractal theory towards human-robot interaction

Abstract

Emotion recognition is crucially related with friendly and humanistic human-robot interaction. Our paper aims at developing a new kind of features to mandarin emotional speech signal based on multifractal theory. Firstly, phase space structure differentiate with respect of initials and finals indicate the fractal phenomenon during speech produce process. Further, positive largest Lyapunov exponent proved existing chaos. To quantitatively measure the chaos, extension of fractal concept-multifractal is calculated by multi-fractal detrended fluctuation analysis (MFDFA) and Legendre transformation. Besides, the underlying fractal characteristics during calculation process is analyzed, which further verifies the emotional speech is multifractal rather than monofractal. Multifractal spectrum visually shows that various emotion is differentiated with each other. After extracting parameters of mulfractal spectrum, several comparative experiments are established, which is implemented with BP neural network and support vector machine (SVM) respectively to hence the comparison between our approach and conventional one. At last, improvement in recognition accuracies demonstrates that our method is available and effective.