Exploiting Annotators' Typed Description of Emotion Perception to Maximize Utilization of Ratings for Speech Emotion Recognition.