Brief Speech Samples Reveal Emotional States in Daily Life

Does our speech – what we say and how we say it – reveal information about our emotional experiences? Scholars have long made assumptions about such links but most evidence is derived from emotion expressions in controlled laboratory work, with only limited studies of emotional experience in naturalistic settings. Using both modern foundation models and traditional speech analysis methods (LIWC and prosodic descriptors), we tested whether brief, prompted speech samples recorded via smartphones could predict concurrent, self-reported emotional states in daily life (N = 934 participants; 12,285 observations). Cross-validated models showed that speech samples predict emotional states (contentment: r_md = .37; sadness: r_md = .23; arousal: r_md = .34), with spoken content captured by foundation-model text embeddings revealing the strongest emotional signal. The findings offer insights into the linguistic characteristics of naturalistic emotional speech, with implications for scalable emotion inference via speech in daily life.