Cognitive load classification using formant features

The ability to automatically classify different cognitive load levels can be very useful, especially in the field of human computer interaction, as human task performance is related to the cognitive load experienced. Although Mel-frequency cepstral coefficients (MFCCs) are commonly used in current speech-based cognitive load classification systems, they offer relatively little insight into how cognitive load affects the speech spectrum and physical speech production system-an area of research which remains poorly understood. Since formants are directly related to the physical characteristics of the vocal tract, we propose the novel use of formant frequencies, bandwidths and formant-based regression coefficients for cognitive load classification. Three-class classification results showed that formant frequencies performed comparably to MFCCs. Additionally, formant frequency-based regression coefficients outperformed MFCC-based regression coefficients by a relative improvement of about 11%. These results imply that formants contain important cognitive load information.