Voice as objective biomarker of stress: association of speech features and cortisol.

Objective: Cortisol is a well-established biomarker of stress, assessed through salivary or blood samples, which are intrusive and time-consuming. Speech, influenced by physiological stress responses, offers a promising non-invasive, real-time alternative for stress detection. This study examined relationships between speech features, state anger, and salivary cortisol using a validated stress-induction paradigm.

Methods: Participants (N = 82) were assigned to cold (n = 43) or warm water (n = 39) groups. Saliva samples and speech recordings were collected before and 20 minutes after the Socially Evaluated Cold Pressor Test (SECPT), alongside State-Trait Anger Expression Inventory (STAXI) ratings. Acoustic features from frequency, energy, spectral, and temporal domains were analysed. Statistical analyses included Wilcoxon tests, correlations, linear mixed models (LMMs), and machine learning (ML) models, adjusting for covariates.

Results: Post-intervention, the cold group showed significantly higher cortisol and state anger. Stress-related speech changes occurred across domains. Alpha ratio decreased and MFCC3 increased post-stress in the cold group, associated with cortisol and robust to sex and baseline levels. Cortisol-speech correlations were significant in the cold group, including sex-specific patterns. LMMs indicated baseline cortisol influenced feature changes, differing by sex. ML models modestly predicted SECPT group membership (AUC = 0.55) and showed moderate accuracy estimating cortisol and STAXI scores, with mean absolute errors corresponding to ∼ 24-38% and ∼16-28% of observed ranges, respectively.

Conclusion: This study demonstrates the potential of speech features as objective stress markers, revealing associations with cortisol and state anger. Speech analysis may offer a valuable, non-invasive tool for assessing stress responses, with notable sex differences in vocal biomarkers.