Evaluation of error- and correlation-based loss functions for multitask learning dimensional speech emotion recognition

Abstract The choice of a loss function is critical part in machine learning. This paper evaluates two different functions commonly used regression-task dimensional speech emotion recognition — error-based and correlation-based functions. We found that using with concordance correlation coefficient (CCC) resulted better performance than mean squared error (MSE) absolute (MAE). evaluations were measured averaged CCC among three emotional attributes. results are consistent input feature sets datasets. scatter plots test prediction by those also confirmed the scores.