Automatic speech analysis can predict loneliness

Abstract

Abstract Loneliness has been demonstrated to exert a detrimental effect on mental and physical health. It may impede the formation of new social relationships by altering interactional behavior. This study provides a proof-of-concept that loneliness is reflected in altered speech, demonstrating that small yet significant effects can make loneliness audible. Samples of 96 healthy participants (mean age 31.08 years, 53 women) were recorded while they performed a picture description and storytelling task. Paralinguistic markers related to prosodic, formant, source, and temporal qualities of speech were extracted and correlated with loneliness, social anxiety and depression. To validate the diagnostic power, machine learning analyses were conducted for women and men separately. A model comprising all speech features from the picture description task significantly predicted loneliness. However, this model did not predict loneliness from the storytelling task. No single speech feature emerged as a strong predictor of loneliness. A combined model that included both speech features and psychiatric symptoms provided better predictions than psychiatric symptoms alone only in women. Overall, these findings suggest that speech offers a new perspective on how loneliness becomes perceptible to others and how it may disrupt social interactions, thereby fostering chronicity.