Abstract:  A method of optimizing feature space for speech emotion recognition is proposed. To achieve better classification between each emotion class; feature space of each pair of emotions were optimized respectively; decomposition of multi-class classifier into two-class classifiers was studied; a decision fusion technique was introduced to re-compose the two-class classifier set; recognition results of multi-class classifier and two-class classifier set were compared in a computer experiment. The results show， recognition rates were improved more than 8 percent under identical environments. The method in this paper， decomposition of multiclass classifier， optimizing feature space of each pair of emotions and decomposition using decision fusion algorithm， is suitable for speech emotion recognition and effective in optimization of feature space.