Improving speech emotion recognition system for a social robot with speaker recognition

摘要

This paper presents modification of a speech emotion recognition system for a social robot. Using speaker dependent classifiers with prior speaker identification step was proposed. Emotion recognition is done using global acoustic features of the speech. Six speech signal parameters are computed with the specialised software. The feature extraction is based on calculation of global statistics of those parameters and their derivatives. Correlation based feature selection algorithms are used for reducing feature vector length. For the evaluation of the system three classifiers were used: Bayes Network, Radial Basis Function Network and Support Vector Machine. Six emotions could be recognised by the system. Several experiments were carried out: emotion recognition of the speaker not known to the system, speaker independent recognition of the known speakers, gender dependent classification as well as speaker dependent recognition with prior identification. The results of identification-driven speech emotion recognition show significant improvement compared to the performance of the speaker independent system based on previous research. The use of the tailored speaker dependent recognition system led to increase in the Bayes Network classification accuracy from 74% to 96%, in the Support Vector Machine classification accuracy from 64% to 96% and in the Radial Basis Function Network classification accuracy from 67% to 88%.