Automatic Detection of Stress from Speech in the Trier Social Stress Test

This study investigates whether speech can serve as an unobtrusive, objective biomarker for acute psychosocial stress. Specifically, the researchers aimed to determine if acoustic-prosodic features extracted from speech could distinguish between individuals undergoing the Trier Social Stress Test (TSST)—a gold-standard stress-induction paradigm—and those in a friendly, non-stressful control condition (f-TSST). Furthermore, the study explored whether these speech features could predict specific physiological markers (cortisol and salivary alpha-amylase) and changes in self-reported affect.

Researchers collected speech data from 50 participants randomly assigned to either the TSST or the f-TSST. The team used a robust processing pipeline involving speaker diarization to isolate participant speech, followed by the extraction of 144 acoustic features, including Mel-Frequency Cepstral Coefficients (MFCCs), classical voice parameters (e.g., pitch, jitter, shimmer), and the eGeMAPS feature set. They trained four machine learning models (Logistic Regression, SVM, Random Forest, and XGBoost) to classify the conditions and used regression models (SVR, Random Forest, and XGBoost) to predict stress reactivity. Model performance was evaluated using nested cross-validation, and feature importance was analyzed using SHAP values.

The XGBoost classifier achieved the highest accuracy (82%) in distinguishing between the stress and control conditions, significantly outperforming the baseline. The most informative features for this classification included the variability of voiced spectral flux, spectral energy at low frequencies, and the rate of voiced speech segments. Additionally, the models successfully predicted cortisol reactivity and changes in negative affect, with performance exceeding that of a mean-baseline regressor. These results suggest that speech contains reliable, measurable markers of the human stress response.