Speech emotion recognition via a max-margin framework incorporating a loss function based on the Watson and Tellegen's emotion model