Brief speech samples reveal emotional states in daily life

Abstract

Does our speech—“what” we say and “how” we say it—reveal how we subjectively feel in daily life? Speech and emotion are often thought to be linked, but most supporting evidence comes from controlled laboratory settings and focuses on expressed, enacted, or externally labeled emotions, leaving open the question of whether naturalistic speech reflects subjective emotional experience in daily life. To address this question, we extracted both modern foundation-model embeddings and established speech-analysis variables from brief, prompted smartphone recordings, including Linguistic Inquiry and Word Count features and prosodic descriptors, and used them as inputs to supervised machine learning models predicting concurrent self-reported emotional states (934 participants; 12,285 observations). Cross-validated models evaluated on unseen participants captured information about self-reported emotional states (contentment: median Spearman ρ = 0.37; sadness: ρ = 0.24; arousal: ρ = 0.35), with spoken content represented by foundation-model text embeddings carrying the strongest emotional signal. Interpretability analyses provided further insights into the linguistic characteristics of everyday emotional language. These findings provide large-scale evidence that brief speech samples contain information about subjective emotional states in daily life.