Cognitive Load Estimation from Speech Commands to Simulated Aircraft

This paper investigates and compares methods for cognitive load (CL) estimation from speech. The majority of previous studies of CL estimation used speech collected in laboratory conditions and conventional speech classification methods. Traditionally laboratory speech contains balanced classes that are labeled by a third party after the speech has been collected. In contrast, the speech used in this research was recorded during an experiment focused on human-machine interaction - where spoken commands were used to control simulated aircraft. The speech was labeled using subjective assessments of CL during an experiment that manipulated workload. Current state-of-the-art Convolutional Neural Network (CNN) classification was used for cognitive load estimation and was compared with conventional Support Vector Machine (SVM) and k-Nearest Neighbor (k-NN) classification. Different speaker-dependence models were compared across 2 and 3 classes. In addition, class boundary selection was optimized to reflect the subjective human workload response sigmoidal curve and compared with linear class boundaries. Results for 3-class CL estimation showed that CNN classifiers trained using speech spectrograms for Partially Speaker Dependent (PSD) models using sigmoidal curve class boundaries provided up to 83.7% accuracy. CNN classifiers outperformed baseline SVM and k-NN classifiers (that used acoustic features) on the same dataset by 13.2% and 10.5% respectively. These outcomes indicate that spectrogram-trained CNN classifiers are a worthy consideration in paralinguistic classification problems.