Voice-based stress detection: A systematic review of machine- and deep-learning approaches (2021–2025)

Stress detection from speech has gained increasing attention as an unobtrusive and scalable approach for real-time health monitoring, supported by advances in machine learning (ML) and deep learning (DL). Because stress-related states may induce measurable alterations in vocal production—including shifts in pitch, intensity, and spectral structure—speech has emerged as a promising source of computational stress biomarkers. This systematic review synthesizes peer-reviewed studies published between 2021 and 2025 that applied ML/DL models to classify stress using audio signals. Following PRISMA 2020 guidelines, 12 studies met the eligibility criteria. Extracted information included dataset characteristics, stress-induction paradigms, acoustic feature sets, model architectures, validation strategies, and reported performance metrics. A descriptive quantitative synthesis was performed, with standard errors and 95% confidence intervals calculated when sufficient statistical information was available. Reported accuracies varied widely, from 39% in multi-class laboratory tasks to above 95% in acted or emotion-labeled corpora. Deep-learning architectures achieved the highest peak accuracies, whereas traditional ML models remained competitive in small-sample laboratory settings. Subgroup analyses revealed systematic differences across dataset types: acted emotional speech yielded the highest performance, laboratory-elicited stress produced moderate accuracy, and real-world recordings showed greater variability. QUADAS-2 assessment indicated notable heterogeneity in dataset realism and reference-standard validity. Overall, current evidence demonstrates the technical feasibility of voice-based stress biomarkers but highlights the need for standardized protocols, physiologically grounded labels, and ecologically valid datasets to support reliable deployment in practical environments.