Predicting cognitive load levels from speech data

An analysis of acoustic features for a ternary cognitive load classification task and an application of a classification boosting method to the same task are presented. The analysis is based on a data set that encompasses a rich array of acoustic features as well as electroglottographic (EGG) data. Supervised and unsupervised methods for identifying constitutive features of the data set are investigated with the ultimate goal of improving prediction. Our experiments show that the different tasks used to elicit the speech for this challenge affect the acoustic features differently in terms of their predictive power and that different feature selection methods might be necessary across these sub-tasks. The sizes of the training sets are also an important factor, as evidenced by the fact that the use of boosting combined with feature selection was enough to bring the unweighted recall scores for the Stroop tasks well above a strong support vector machine baseline.