Automatic cognitive load detection from speech features

Abstract

Cognitive load variations have been found to impact multimodal behaviour, in particular, features of spoken input. In this paper, we present a design and implementation of a user study aimed at soliciting natural speech at three different levels of cognitive load. Some of the speech data produced was then used to train a number of models to automatically detect cognitive load. We describe a classification approach, the cognitive load levels were detected and output as discrete level ranges. The final system achieved a 71.1% accuracy for 3 levels classification in a speaker-independent setting. The ability to detect and manage a user's cognitive load can help us to adapt intelligent interfaces that ensure optimal user performance