Continuous affect prediction using eye gaze and speech

Abstract

Affective computing research traditionally focused on labeling a person's emotion as one of a discrete number of classes e.g. happy or sad. In recent times, more attention has been given to continuous affect prediction across dimensions in the emotional space, e.g. arousal and valence. Continuous affect prediction is the task of predicting a numerical value for different emotion dimensions. The application of continuous affect prediction is powerful in domains involving real-time audio-visual communications which could include remote or assistive technologies for psychological assessment of subjects. Modalities used for continuous affect prediction may include speech, facial expressions and physiological responses. As opposed to single modality analysis, the research community have combined multiple modalities to improve the accuracy of continuous affect prediction. In this context, this paper investigates a continuous affect prediction system using the novel combination of speech and eye gaze. A new eye gaze feature set is proposed. This novel approach uses open source software for real-time affect prediction in audio-visual communication environments. A unique advantage of the human-computer interface used here is that it does not require the subject to wear specialized and expensive eye-tracking headsets or intrusive devices. The results indicate that the combination of speech and eye gaze improves arousal prediction by 3.5% and valence prediction by 19.5% compared to using speech alone.

Eye-based Continuous Affect Prediction

“…From the results in [22] the bimodal fusion of EEG and eye-based features performed best overall (arousal = 67.7%, valence = 76.1%). Eye gaze was combined with speech in [28], where a feature set similar to that of [22] was used. Additional statistics were gathered for eye scan paths and eye closure features were measured by frame counts instead of time.…”

“…Additional statistics were gathered for eye scan paths and eye closure features were measured by frame counts instead of time. Results achieved in [28] showed that eye gaze, when combined with speech as part of a feature fusion, single support vector regression system, could improve arousal prediction compared to that of unimodal speech (3.5% relative performance improvement), while model fusion improved valence prediction compared to unimodal speech (19.5% rela-tive performance improvement). Psychopathological affective computing work incorporating eye-based features as part of multimodal approaches include post traumatic stress disorder estimation [29] and depression recognition [30], [31].…”