A hierarchical support vector machine based on feature-driven method for speech emotion recognition

Through the analysis of one-vs.-one, one-vs.-rest and the decision tree mechanism of binary support vector machine emotion classifiers, a method based on feature-driven hierarchical support vector machine is proposed for speech emotion recognition. For each layer, classifier used different feature parameters to drive its performance, and each emotion is subdivided layer by layer. This method did not rely entirely on the activity-valance dimensional emotion model, but relied on the type of emotion to distinguish. Furthemore, classifications are constructed by appropriate characteristic parameters ultimately. Experiments on the Chinese-speakerdependent and Berlin-speaker-independent corpus reached conclusions as follows, Chinese-speaker-dependent recognition rate is relatively higher than Berlin-speaker-independent. feature-driven hierarchical support vector machine in the case driven by effective features improves the speech emotion recognition performance. Meanwhile applying the mean of the log-spectrum to this method can identify high-activity and lowactivity emotion effectively.