Predicting affective engagement and mental strain from prosodic speech features.

Background: Emotional resilience (traditionally defined as the capacity to recover from adversity) and cognitive load (the mental effort for processing information) are critical aspects of mental health functioning. Traditional assessment methods, such as physiological sensors and post-task surveys, often disrupt natural behavior and fail to provide real-time insights. Speech prosody, encompassing pitch, intensity, loudness, and voice activity, offer a non-intrusive alternative for evaluating these psychological constructs. However, the relationship between speech prosody, emotional resilience, and cognitive load remains underexplored, particularly in conversational contexts. Objective: This study proposes proxy measures for these constructs based on self-reported engagement, enjoyment, boredom, and cognitive effort during dyadic conversation. By leveraging the SEWA (Automatic Sentiment Estimation in the Wild) database, developed through a European research project on emotion recognition, the research seeks to develop machine learning models that correlate speech patterns with subjective self-reports of emotional and cognitive states. Methods: Prosodic features, such as pitch variation, vocal intensity, and voice activity, were extracted from the SEWA database recordings. These features are then normalized to account for inter-speaker variability and used as predictors in machine learning models. Regression and classification models are employed to correlate speech features with subjective self-reports, which serve as ground truth for Positive Affective Engagement (as a proxy for emotional resilience) and Perceived Mental Strain (as a proxy for cognitive load). Data from English and German speakers are analyzed separately to account for linguistic and cultural differences. Outcomes: The study establishes a significant relationship between speech prosody and psychological states, demonstrating that Positive Affective Engagement (as a proxy for emotional resilience) and Perceived Mental Strain (as a proxy for cognitive load) can be effectively predicted through prosodic features. Higher emotional resilience is linked to more discernible prosodic patterns in German speech, such as higher loudness and greater voice probability consistency. In contrast, cognitive load prediction remains consistent across English and German datasets. Conclusion: This research introduces a novel approach for assessing Positive Affective Engagement (as a proxy for emotional resilience) and Perceived Mental Strain (as a proxy for cognitive load) through speech prosody, highlighting the significant impact of language-specific variations. By combining prosodic features with machine learning techniques, the study offers a promising alternative to traditional psychological assessments. The findings emphasize the need for tailored, multilingual models to accurately estimate psychological states, with potential applications in mental health monitoring, cognitive workload analysis, and human-computer interaction. This work lays the foundation for future innovations in speech-based psychological profiling, advancing our understanding of human emotional and cognitive states in diverse linguistic contexts.