Two-Dimensional Cepstrum Analysis Approach in Emotion Recognition from Speech

In this paper we propose two-dimensional cepstrum analysis approach applied in the emotion recognition task in continuous space. Experiments has been done on IEMOCAP database which allows emotion recognition in three dimensions: dimension valence, arousal and dominance. Sequence of experiments has been done over whole range of emotions available in the IEMOCAP database. It includes 6 basic emotional states with three additional emotional states. Performance of system based on two-dimensional cepstrum approach is immediate compared with different approaches. For that cause we utilized concept of "bag-of-features". It includes vast collection of features with final amount of 961 features per vector. Then we selected most appropriate features with Bat Algorithm and Particle Swarm Optimization method. Performance of our system has been evaluated with Pearson correlation coefficient. Results from both systems presented in this contribution show that we achieved better results in dimension arousal (DA), dimension dominance (DO) and similar results in last one dimension valence (DV).