A novel objective function to optimize neural networks for emotion recognition from speech patterns

Expressions of different emotions are usually overlapping and hard to distinguish. Besides, the amounts of feature patterns are usually imbalanced in the overlapping emotional expressions, and most conventional classifiers tend to prefer lager classes for archiving a better overall recognition rate. This drawback is also encountered in the Multi-layer perception (MLP) models frequently proposed for emotion recognition due to its superior classification capability and performance. However, MLP and most recognition techniques only refer to a mean square error and an overall error rate. Furthermore, using MLP has another disadvantage that needs to search a suitable network structure. In this paper, a novel objective function to optimize the MLP neural networks is proposed for solving these problems. This function considers the criteria of mean square error, classification error rate, and distances between the examples and the classification boundary, to optimize the network parameters and prune the links between neurons. Besides, the sigmoid and Gaussian transfer functions are adopted in our method to construct suitable classification boundaries. An artificial data set and the Danish emotional speech database are used to verify the MLP-based classifier with the novel objective function. The experimental results show that the proposed model has better performance than conventional MLPs.