Significant outcomes
-
• Speech features showed clear differences between cold and warm conditions, reflecting stress-related changes.
-
• Alpha ratio and MFCC3 emerged as the most reliable speech markers of stress, capturing involuntary physiological vocal changes.
-
• Local shimmer also indicates stress through reduced amplitude variability, while frequency-based features like jitter and pitch are less reliable due to sex and individual variability.
Limitations
-
• Variability in technical equipment across recordings may affect the comparability of acoustic features.
-
• The study did not control for participants’ circadian rhythms, which can influence cortisol measurements.
-
• The SECPT stress induction paradigm may not fully represent the complexity of real-world stress responses.
-
• The study lacks a recovery-phase cortisol measure, limiting interpretation of whether speech changes reflect peak stress or recovery.
Introduction
Stress is a complex physiological and psychological response to external and internal stimuli. A key biological marker of stress is cortisol, a glucocorticoid hormone released by the adrenal cortex (Grant et al., Reference Grant, Forrest and Symington1957). Cortisol plays a crucial role in stress response, impacting metabolism, immune response, and central nervous system activity (Marieb, Reference Marieb and Hoehn2007). Elevated cortisol levels are associated with heightened stress, making it a reliable biomarker for stress assessment (Dickerson & Kemeny, Reference Dickerson and Kemeny2004; Hellhammer et al., Reference Hellhammer, Wüst and Kudielka2009).
Stress is a significant risk factor for poor health (O’Connor et al., Reference O’Connor, Thayer and Vedhara2021), including depression and anxiety disorders (Hammen, Reference Hammen2015; Pêgo et al., Reference Pêgo, Sousa, Almeida, Sousa, Stein and Steckler2010).Underestimating stress is associated with long-term development of depressive symptoms (Izawa et al., Reference Izawa, Nakamura-Taira and Yamada2016). Given these risks, objective and accurate stress measurement is crucial for better management and improved outcomes under pressure (Liu et al., Reference Liu, Ein, Gervasio, Vickers and Zhou2019). However, traditional methods like self-report questionnaires are subjective and susceptible to bias. This underscores the need for objective, reliable, and non-invasive stress biomarkers (Slavich, Reference Slavich2019).
Multimodal wearable devices studies have shown that combinations of physiological signals (e.g., heart rate variability, electrodermal activity, electroencephalography, blood-volume pulse, inter-beat interval, skin temperature) can distinguish stress states and track changes referenced to salivary cortisol (Betti et al., Reference Betti, Lova, Rovini, Acerbi, Santarelli, Cabiati, Del Ry and Cavallo2018; Nath & Thapliyal, Reference Nath and Thapliyal2021). Another promising measure for stress biomarker development might be speech, as a natural and ubiquitous mode of communication. A direct link between physical state and speech production has been proposed several decades ago (Scherer, Reference Scherer1986). Since stress increases muscle tension and respiration rate, which subsequently affect speech production, research suggests stress could be identifiable through the characteristics of speech (Sondhi et al., Reference Sondhi, Khan, Vijay and K. Salhan2015). These include speech fluency (Buchanan et al., Reference Buchanan, Laures-Gore and Duff2014), fundamental frequency, i.e. pitch (Rajasekaran et al., Reference Rajasekaran, Doddington and Picone1986), jitter, speaking rate, and length and number of pauses (J. H. L. Hansen & Patil, Reference Hansen, Patil and Müller2007a). Recent findings further highlight stress effects on features like fundamental frequency (F0), Harmonics-to-Noise Ratio (HNR), and shimmer (Kappen et al., Reference Kappen, van der Donckt, Vanhollebeke, Allaert, Degraeve, Madhu, Van Hoecke and Vanderhasselt2022).
Research shows significant sex differences in voice quality under stress. Women tend to experience more vocal symptoms and stress symptoms than men (Holmqvist et al., Reference Holmqvist, Santtila, Lindström, Sala and Simberg2013). Under stress, female voices become breathier, more strained, and exhibit lower fundamental frequency, intensity, and aerodynamic capacity (Van Lierde et al., Reference Van Lierde, Van Heule, De Ley, Mertens and Claeys2009). For high-anxiety females, articulation precision increases under cognitive stress but decreases under emotional stress (Tolkmitt & Scherer, Reference Tolkmitt and Scherer1986). Stress-induced changes in muscle tension can lead to increased F0 floor in both high-anxious and anxiety-denying individuals (Tolkmitt & Scherer, Reference Tolkmitt and Scherer1986). These findings suggest stress affects vocal quality differently by sex, with women showing more pronounced changes.
Using speech to detect stress offers advantages like continuous, real-time, and non-invasive monitoring (Robin et al., Reference Robin, Harrison, Kaufman, Rudzicz, Simpson and Yancheva2020), making it ideal in environments where traditional methods are impractical. In telemedicine and remote patient monitoring settings, speech analysis could provide immediate feedback on patient stress levels without requiring physical presence or complex equipment (König et al., Reference König, Riviere, Linz, Elbaum, Fabre and Robert2021).
The link between speech features and cortisol, a validated physiological stress marker, remains underexplored. Pisanski and Sorokowski (Reference Pisanski and Sorokowski2021) found that elevated salivary cortisol levels in 10 female students during a real-life stressful situation correlated with increased pitch, altered vocal tract resonances, and changes in speech speed, leading listeners to perceive these voices as more stressed. Baird and colleagues (Reference Baird, Triantafyllopoulos, Zänkert, Ottl, Christ, Stappen, Konzok, Sturmbauer, Meßner, Kudielka, Rohleder, Baumeister and Schuller2021) used a multi-modal approach and identified significant speech features, including pitch, speech rate, and formant frequencies as effective predictors of physiological stress markers including saliva-based cortisol levels, heart rate, and respiration. Other multimodal studies using biosignals and machine learning have achieved up to 95.21% accuracy in classifying stress states (Aigrain et al., Reference Aigrain, Spodenkiewicz, Dubuisson, Detyniecki, Cohen and Chetouani2018; Bobade & Vani, Reference Bobade and Vani2020). Despite these findings, gaps remain in understanding how specific speech features correlate with cortisol fluctuations across various contexts and populations.
Combining speech analysis with cortisol measurement offers a promising approach to stress detection, paving the way for precise diagnostic tools. Validating speech-based biomarkers against objective stress indicators like cortisol is essential to ensure their reliability and utility, particularly for early diagnosis of stress-related disorders or real-time monitoring in high-pressure professions. This study investigates the associations between speech features and cortisol levels to evaluate the potential of speech as an objective stress biomarker. Based on existing literature, we posit that features such as increased voice pitch, altered vocal tract resonances, voice quality features, and changes in speech speed will correlate with elevated cortisol levels following the stress-inducing paradigm. Due to inconsistent findings, the direction of these effects remains uncertain.
Our primary aim was to test whether acute stress induced by the Socially Evaluated Cold Pressor Test (SECPT) produces vocal changes that covary with hypothalamic–pituitary–adrenal activation indexed by salivary cortisol reactivity measured 20 minutes after stressor onset. We examined whether a set of vocal features reflects acute stress when anchored to a physiological marker of HPA activation. Specifically, we tested whether changes in selected acoustic parameters from pre- to post-SECPT relate to salivary cortisol reactivity and whether these relationships differ between sexes.
Material & methods
Participants
Healthy participants were recruited for this study over a 15-month period at the Department of Psychiatry, University of Frankfurt, Germany. The data reported here stem from two separate experiments within the larger project ‘Identifying Mediators of Stress-Aggression and Experimental Manipulation by tDCS’, approved by the Ethics Committee of the Goethe University Frankfurt Medical Faculty (Ref: 20-1035). The study adhered to the Declaration of Helsinki. All participants provided written informed consent and received €10 per hour for their participation.
Inclusion criteria required participants to be right-handed, aged 18–55, German language proficiency at C1 level, with no confirmed psychiatric, neurological, or cardiovascular conditions. To exclude cognitive impairments and low IQ (<70), we administered the Trail-Making Test (Reitan, Reference Reitan1955) and the Mehrfachwahl-Wortschatz-Intelligenztest (MWT-B, Lehrl, Reference Lehrl2005). Additional exclusion criteria included smoking (more than 10 cigarettes/day), caffeine intake above 400 mg/day, adipositas, and any medication with potential effects on the HPA axis activity, e.g., current use of cortisone nasal spray. Participants were instructed to abstain from alcohol and other drugs for 48 hours prior to testing. If any participant showed signs of a clinically significant psychiatric or somatic condition during the assessment, they were advised to seek appropriate clinical care.
Study overview
This study employed a randomised 2 × 2 factorial design with time point (pre-stressor vs. post-stressor) as the within-subject factor and group (cold water vs. warm water) as the between-subject factor. All procedures were conducted at the Department of Psychiatry, Psychosomatics, and Psychotherapy, Goethe University Hospital Frankfurt, Germany. To control for cortisol’s diurnal variation, all experimental sessions were scheduled between 2:00 PM and 6:00 PM (Lee et al., Reference Lee, Kim and Choi2015). Fig. 1 provides an overview of the study timeline.
Graphical overview depicting the study timeline.
Upon arrival, participants received detailed study information. Subsequently, the experimenter documented recent behaviours affecting cortisol levels, including food intake, smoking, caffeine consumption, gum chewing, tooth brushing, and physical activity (Fukuda & Morimoto, Reference Fukuda and Morimoto2001; Pritchard et al., Reference Pritchard, Stanton, Lord, Petocz and Pepping2017) and verified compliance with pre-experimental restrictions. To minimise confounders, subjects were not allowed to eat or drink anything other than water for one hour before the testing session started..
Baseline stress assessments included the first salivary cortisol sample (see cortisol assessment), completion of the State–Trait Anger Expression Inventory (STAXI), and a speech recording (see Psychological assessment and speech recording). Afterwards stress induction was achieved using the Socially Evaluated Cold Pressor Test (SECPT, see Stress test procedure), followed by a second STAXI and speech recording. The second cortisol sample was obtained 20 minutes after SECPT onset. For half the participants, the same procedure was followed with an added EEG recording, not included in this analysis..
Stress test procedure
To induce stress, we used the Socially Evaluated Cold Pressor Test (SECPT), a well-established paradigm combining physical discomfort and social evaluation to elicit psychosocial stress (Schwabe & Schächinger, Reference Schwabe and Schächinger2018). The cold water triggers a physiological stress response by activating nociceptors, sending signals to the hypothalamus, the primary stress regulator. Simultaneous introduces a psychosocial stressor, further amplifying the response. This combination is particularly effective in activating the Hypothalamic-Pituitary-Adrenal (HPA) axis, as shown by elevated salivary cortisol levels.
Participants were randomly assigned to a cold or warm water group. In the cold water group (N = 43), participants immersed their right hand up to the wrist in 3–5°C water. They were informed that the session would be video-recorded to study their reactions (Schwabe et al., Reference Schwabe, Haddad and Schachinger2008) asked to look into a camera, and instructed to keep their hand submerged for as long as possible, up to 3 minutes. A researcher observed them throughout. No actual filming took place, which was disclosed during debriefing. Participants in the warm water group (N = 39) immersed their hands in 35–37°C warm water for 3 minutes without video-recording.
Psychological assessment and speech recording
To evaluate psychological stress, participants self-reported anger symptoms before and after the intervention using the State–Trait Anger Expression Inventory (STAXI, Spielberger et al., Reference Spielberger, Sydeman, Owen and Marsh1999). Participants also completed a speech task, reading 20 short sentences (e.g., ‘The rag is on the freezer’. ‘They just carried it upstairs and now they’re going down again’.) both pre- and post-intervention. A full sentence list is provided in the supplementary material. Responses were recorded using different mobile phones and stored in .wav-Format.
Cortisol assessment
Two saliva samples were collected from all participants in each session: the first after completing the initial questionnaires, and the second 20 minutes after SECPT onset. Saliva was collected using the ‘Salivette’ (Sarstedt AG & Co., Nümbrecht, Germany) which includes a cotton swab suspended in a centrifuge vessel. Participants gently chewed the cotton swab for one minute, then returned it to the insert without touching. Salivary samples were stored at −30°C until the biochemical analysis, which was performed by the Group of Translational Psychiatry, University Hospital Frankfurt, Germany. Samples were stored for a maximum of 47 days until analyses. The concentration of free salivary cortisol was analysed using a luminescence immunoassay (IBL, Hamburg, Germany) with intra- and inter-assay precision of 4.5% and 4.3%, respectively.
Speech feature processing
Automatised acoustic feature extraction was conducted with our own speech processing library ‘Sigma‘ using Python 3.9. Scripts are available upon reasonable request from the corresponding author.
We investigated several categories of acoustic speech features previously associated with stress symptoms (Buchanan et al., Reference Buchanan, Laures-Gore and Duff2014; J. Hansen & Patil, Reference Hansen, Patil and Müller2007a; Sondhi et al., Reference Sondhi, Khan, Vijay and K. Salhan2015). Attributes were categorised into groups (refer to Supplementary Table 1 for groupings and corresponding attributes), including frequency, energy, spectral and temporal features. Frequency features consist of variables associated with the formants F0 to F3. F0 is perceived as voice pitch (Ladefoged, Reference Ladefoged1996), while formants F1-F3 contribute to vowel sound articulation (Ladefoged & Johnson, Reference Ladefoged and Johnson2011). Energy features included jitter(voice instability), shimmer (variability in loudness or intensity), and the harmonics-to-noise ratio (HNR (Teixeira et al., Reference Teixeira, Oliveira and Lopes2013)). Spectral features describe frequency characteristics, such as the mel frequency cepstral coefficient (MFCC), often used in speaker identification (Nakagawa et al., Reference Nakagawa, Asakawa and Wang2007). Additionally, we employed features defined by our research team concerning the temporal dimensions of speech (König et al., Reference König, Linz, Zeghari, Klinge, Tröger, Alexandersson and Robert2019), capturing timing aspects of speech, such as duration, rhythm, and temporal patterns (Zellner, Reference Zellner1994). In total 65 features were extracted; 10 energy-related features, 23 frequency-related features, 22 spectral features, and 10 temporal features.
Statistical analysis
Power analysis
The required sample size was determined using G*Power(Faul et al., Reference Faul, Erdfelder, Lang and Buchner2007) with α = 0.05 (one-tailed), β = .80. A small (.25) effect size with two groups (cold/warm water) and two measurements (pre/post measurement) in a repeated measure ANOVA with within and between factor interaction was used in order to calculate the sample size. A correlation of 0.57 between repeated measures, based on the internal consistency of the STAXI-II subscales (Schamborg et al., Reference Schamborg, Tully and Browne2016) informed the calculation. To allow for separate analysis by sex, a total of N = 60 participants was required.
Descriptive statistics
Statistical analyses were conducted with the Python package scipy.stats (v1.11.4, Linux v5.10.0). Descriptive statistics were computed for demographic variables, cortisol, and STAXI values.
Speech features
Within-group differences in speech features pre- and post-intervention were assessed using the paired Wilcoxon test, a non-parametric method suitable for small samples or non-normal data (Wilcoxon, Reference Wilcoxon1945).
Each of the twenty sentences was analysed independently. Sentence order effects were controlled by including sentence ID as a covariate. P values were adjusted for multiple hypothesis testing using the Benjamini–Hochberg procedure (Benjamini & Hochberg, Reference Benjamini and Hochberg1995) clustered in categories (see Supplementary Table 1).
To assess associations between speech features and cortisol/STAXI values while accounting for baseline differences, we computed changes scores (post–pre) for each measure. Spearman correlations were then calculated between these difference scores. Sex-based differences in correlation strength were assessed using Fisher’s Z transformation to compare correlation coefficients between male and female participants.
Linear models (LMM) were applied to examine the relationship between cortisol changes and speech feature changes, accounting for potential moderating effects of group, sex, and baseline cortisol levels. The model used was: Cortisol_Change ∼ Group * Feature_Change + Sex + Cortisol_Before. LMMs accounted for repeated measures by including random effects for participant-specific intercepts and slopes, while fixed effects modelled relationships between speech features and cortisol change. Outliers were removed using the interquartile range (IQR) method (Vinutha et al., Reference Vinutha, Poornima, Sagar, Satapathy, Tavares, Bhateja and Mohanty2018), applied independently by sex and experimental group for each speech feature. All change scores were calculated as post-intervention minus pre-intervention values.
To evaluate the predictive value of speech features, machine learning models were trained to (1) classify group membership (cold vs. warm) and (2) predict cortisol levels and STAXI scores. To control sex-related variability, each acoustic feature was regressed on sex, and the residuals were used as sex-normalised inputs. All features were Z-score standardised. Models were trained on the difference scores (after - before) to capture intervention effects. Five classifier algorithms were evaluated: support vector machines (SVM), extra trees (XT), random forests (RF), linear models (LM), and decision trees (DT). Model performance was assessed using 10-fold group cross-validations, with folds stratified by participant ID to prevent data leakage. Stepwise feature selection was applied: models were iteratively trained using an increasing number of top-ranked features to determine the optimal set. Classification performance was assessed using the area under the ROC curve (AUC), while regression performance was evaluated using mean absolute error (MAE).
Results
A total of 82 participants were included, each undergoing repeated measures at two time points: before and after exposure to the experimental conditions (cold/warm water). Table 1 demonstrates STAXI scores and cortisol levels for both groups (cold/warm) and sexes.
Demographics, STAXI scores and cortisol data
There were no statistically significant differences in STAXI scores between sexes at either time point, based on Wilcoxon rank-sum test (t0 : W = 730, p = 0.247; t0 : W = 802, p = 0.693). Similarly, Cortisol levels did not differ significantly between sexes (t0 : W = 693, p = 0.174; t0 : W = 738, p = 0.346).
Age also did not differ significantly between cold and warm water groups (W = 851, p = 0.911), suggesting that demographic and baseline characteristics were comparable across the conditions and sexes.
Cortisol and STAXI state of anger levels changed significantly pre- to post intervention. In the cold water condition, participants exhibited a significant increase in cortisol (W = 58,100, p < 0.001) and STAXI scores (W = 23,390, p < 0.001) after SECPT. Conversely, participants in the warm water condition showed a significant decrease in cortisol (W = 26,680, p < 0.001) and STAXI scores (W = 7,580, p < 0.001).
When analysed by sex, similar patterns emerged. Among females, cortisol (W = 6,630, p < 0.001) and STAXI scores (W = 1,590, p < 0.001) significantly decreased after warm water exposure, whereas both cortisol (W = 1,787, p < 0.001) and STAXI scores (W = 5,430, p < 0.001) significantly increased after cold water exposure. Among males, cortisol (W = 6,630, p < 0.001) and STAXI scores (W = 107.8, p < 0.001) significantly decreased after exposure to the warm water condition. In contrast, cortisol (W = 11,850, p < 0.001) and STAXI scores (W = 6,390, p < 0.001) significantly increased after cold water exposure.
Differences in Speech Features before and after Intervention
A Wilcoxon test was conducted to compare pre- and post SECPT timepoints for both conditions (cold and warm, see Table 2). Reported p values reflect these comparisons..
Difference in speech features before and after intervention. Features that remained significant after correction are highlighted in bold. Before and after arrows indicate if the feature value increased or decreased in value after exposure to cold/warm water. Features are organised by feature type and then alphabetically. Only features that were significant in the warm and/or cold group are summarised in Table 2. All results are listed in Supplementary Table 2
Energy features revealed distinct patterns between groups. Loudness rate increased in both groups. However, while mean loudness decreased in the warm group (n.s.), it exhibited a significant increase in the cold group. The mean harmonic-to-noise ratio (HNR) showed a significant increase in the warm group, accompanied by a decrease in HNR standard deviation. The cold group showed similar trends, though not statistically significant.
Frequency features revealed opposing trends. In the warm group, both the mean and standard deviation (SD) of F1 bandwidth significantly increased, while both decreased in the cold group (n.s.). The mean F2 frequency significantly increased in the warm group and decreased (n.s.) in the cold group. Conversely, the SD of F2 frequency significantly decreased in the cold group, and increased (n.s.) in the warm group. The mean F3 frequency increased in the warm group, and F3 bandwidth increased in the cold group. Mean pitch significantly decreased in the warm group, whereas the cold group exhibited a non-significant opposite trend. Pitch minimum decreased and pitch SD increased in the warm group, with the same pattern observed in the cold group but without statistical significance. Overall, formant features displayed opposing trends, but were rarely significant in both groups for the same features.
The analysis of spectral features further supported group-specific trends. Mel-frequency cepstral coefficients (MFCC) 1, significantly decreased in the warm group but showed an non-significant increase in the cold group. MFCC2 significantly decreased in both groups. MFCC3 exhibited significant opposing patterns, increasing in the warm group and decreasing in the cold group. MFCC4 significantly decreased in the cold group, with the same pattern observed in the warm group but without statistical significance. Regarding relative mean energy, F1 relative mean energy significantly decreased in the warm group and showed a non-significant decrease in the cold group. F3 relative mean energy significantly decreased in the cold group, whereas the warm group displayed an opposite but non-significant trend. The mean H1_a3 difference significantly increased in the warm group but significantly decreased in the cold group. Additionally, the H1_a3 difference SD and the mean H1_H2 difference significantly decreased in the cold group, with the same non-significant patterns observed in the warm group. The harmonic difference H1_H2 SD significantly decreased in both groups. The Hammarberg index significantly decreased in the warm group, whereas the cold group exhibited a non-significant increase. The spectral slope within the 0–500 Hz and 500–1500 Hz ranges decreased in both the warm and cold groups.
Finally, the analysis of temporal features indicated significant changes after exposure across several measures. Duration, pause duration, and the total utterance time significantly decreased following exposure. Notably, all but one (mean utterance duration) of the temporal features exhibited the same pattern in both the warm and cold groups.
Speech feature correlations with cortisol levels and STAXI scores
Cortisol levels
Spearman correlations between each feature and cortisol levels (Table 3) and STAXI scores (Supplementary Table 4) were calculated using the after-before difference. Only significant results are presented in Fig. 2 and Table 3, with full results in Supplementary Table 3.
Delta in correlations of speech features and cortisol differences (after-before) for cold and warm groups for the whole sample and stratified by sex. For brevity reasons, only significant results are displayed here, the full results are displayed in Supplementary Table 3.
Associations of acoustic features with cortisol levels across conditions. The table presents correlation coefficients (corr), effect sizes (ES), and adjusted p-values (p-adj) for the overall population. Sex-specific correlations are indicated separately for males (M) and females (F), with positive ( + ) or negative ( − ) directions. Fisher’s Z-test results, including Z-values and corresponding p-values, are reported under the ‘Sex Comparison’ header to assess differences between sexes. Results are shown for both cold and warm conditions. For brevity reasons, only significant results are reported here. The full results are displayed in Supplementary Table 3
For energy features, mean loudness was significantly correlated with cortisol levels in the cold condition (r = −0.18, p <0.001), while the warm condition showed a weaker and non-significant correlation (r = −0.01, p = 0.91).
Within frequency features, all jitter measures (DDP jitter (r = −0.10, p = 0.03), local absolute jitter (r = −0.10, p = 0.02), PPQ5 jitter (r = −0.10, p = 0.02), and RAP jitter (r = −0.10, p = 0.03) were significantly negatively correlated with cortisol levels in the cold condition. These relationships were weaker and non-significant in the warm group (p >0.08).
For spectral features, the mean alpha ratio showed a small but significant negative correlation with cortisol levels in the cold condition (r = −0.12, p = 0.01), while no notable association was found in the warm condition (r = −0.03, p = 0.77). MFCC1 was negatively correlated with cortisol in the cold condition (r = −0.16, p<0.001) but not significantly associated with the warm condition (r = 0.10, p = 0.09). In contrast, MFCC2 exhibited a significant negative correlation in the warm condition (r = −0.16, p <0.001) but not in the cold condition (r = 0.03, p = 0.57). MFCC4 displayed opposite patterns across conditions, with a weak positive correlation in the cold condition (r = 0.09, p = 0.03) and a weak negative correlation in the warm condition (r = −0.07, p = 0.3), with a significant sex difference (Z = −2.68, p = 0.01).
Regarding formant energy measures, F1 relative energy SD showed a significant negative correlation with cortisol in the cold condition (r = −0.10, p = 0.02), but not in the warm group (r = −0.06, p = 0.3). Mean F3 relative energy was positively correlated with cortisol in the cold condition (r = 0.11, p = 0.01) but not in the warm condition (r = 0.0, p = 1.0).
The Hammarberg index demonstrated a significant negative correlation with cortisol levels in the cold condition (r = −0.12, p = 0.01), with a strong sex effect (Z = 2.89, p <0.001). In the warm condition, this relationship was weak and not significant (r = 0.02, p = 0.9).
STAXI scores
The relationship between acoustic features and STAXI scores varied across the cold and warm stress conditions. Only significant results are presented in Fig. 3, full results are displayed in Supplementary Table 4.
Delta in correlations of speech features and STAXI scores (after-before) for cold and warm groups for the whole sample and stratified by sex. For brevity reasons, only significant results are displayed here, the full results are displayed in Supplementary Table 4.
For energy features, HNR was negatively but not significantly correlated with STAXI scores in the cold condition (r = −0.09, p = 0.08), while in the warm condition, it was significantly positively correlated (r = 0.14, p < 0.001). No significant sex differences were observed in either condition.
Regarding frequency features, PPQ5 jitter was not significantly correlated with STAXI scores in the cold condition (r = −0.04, p = 0.48). However, in the warm condition, it showed a significant positive correlation (r = 0.14, p < 0.001). No significant sex differences were found.
For spectral features, the SD of the H1-H2 harmonic difference was significantly positively correlated with STAXI scores in the cold condition (r = 0.10, p = 0.05), but not in the warm condition (r = 0.02, p = 0.81), with no significant sex differences. The Hammarberg index was significantly negatively correlated with STAXI scores in the cold condition (r = −0.13, p < 0.001), with a significant sex difference (Z = 4.62, p < 0.001). In the warm condition, the correlation was positive but not significant (r = 0.09, p = 0.12), with no significant sex effect.
Linear mixed model analyses
Linear mixed model analyses examining the relationships between cortisol changes, baseline cortisol levels, sex, and changes in speech features revealed several significant findings. Higher baseline cortisol levels were linked to changes in average_mfccs_3 (β = −0.61, p = 0.014) and local_shimmer (β = −0.20, p = 0.003). Changes in cortisol (after - before) were negatively associated with changes in mean alpha ratio (β = −0.05, p = 0.02), indicating that larger cortisol changes were linked to smaller changes in this feature.
Sex differences were also observed: male participants exhibited consistently negative associations with changes in multiple speech features, including features mirroring shimmer, jitter, and pitch. Supplementary Table 5 displays the detailed results.
Machine learning model performance
The highest AUC was achieved by the SVM classifier using difference-based features (AUC = 0.55, k = 15), followed closely by XT (AUC = 0.54) and RF (AUC = 0.53). For regression tasks, SVMs also yielded the best performance in predicting both cortisol and state anger scores across difference-based input. MAEs for cortisol prediction ranged from approximately 3.8 to 6.0 nmol/L, while those for STAXI scores ranged from ∼ 1.3 to 2.2. Full model metrics, including performance across all time points and model types, are presented in Supplementary Table 6. The most frequently selected features across tasks were related to spectral shape and shimmer variability, particularly MFCC 2 and 3s, and voice quality metrics (e.g., apq5_shimmer, f1_bandwidth_mean). Supplementary Table 7 presents the ten most frequently selected features across models.
To support interpretation, a summary table was created (Table 4), consolidating key findings across different analyses. It highlights features significantly affected by sex, linked to stress responses, and robust to potential confounders.
Summary of speech feature associations with sex, stress, and robustness to confounding factors. The table categorises key speech features based on their sensitivity to sex differences (✔ = significant, ✘ = not significant), their link to stress responses (as measured by cortisol levels), and whether these associations are robust to confounding factors such as baseline cortisol and sex. The significant Before v. After column indicates whether a feature showed significant differences before and after exposure to the stress condition and if so, in which condition. Cortisol correlation direction shows whether the feature has a positive ( + ) or negative ( − ) correlation with cortisol levels. Direction (before-after stress) indicates whether the feature difference increased (↑) or decreased (↓) after exposure to the stress condition
Discussion
This study explored whether acoustic speech changes could serve as objective stress markers by linking them to cortisol levels and self-reported anger under induced stress. The SECPT effectively elicited a stress response, evidenced by significant increases in cortisol levels and STAXI state anger scores in the cold water condition. In contrast, the warm water condition produced significant decreases in both measures. These findings align with previous research (Schwabe & Schächinger, Reference Schwabe and Schächinger2018; Hellhammer et al., Reference Hellhammer, Wüst and Kudielka2009) demonstrating the sensitivity of cortisol and anger ratings to acute stressors. Speech features varied distinctly between cold and warm conditions, with spectral features demonstrating the greatest potential as stress markers due to their robustness against confounding variables. This suggests that HPA axis reactivity may influence vocal spectral properties. These findings support existing evidence that stress affects speech production through e autonomic nervous system activation, increasing laryngeal muscle tension and altering respiratory patterns, which, in turn, alter speech energy, frequency, and spectral characteristics (Giddens et al., Reference Giddens, Barron, Byrd-Craven, Clark and Winter2013).
Voice changes under stress stem from a combination of physical and physiological mechanisms, both voluntary and involuntary. However, not all speech features respond uniformly. Individual variability, such as sex and baseline cortisol, complicates interpretation, emphasising the need for robust, stable acoustic indicators. (Hansen & Patil, Reference Hansen, Patil and Müller2007a).
Dahl & Stepp (Reference Dahl and Stepp2023) identified spectral and energy-based measures as more reliable than frequency-based features due to greater consistency across sexes. We identified alpha ratio and MFCC3 as especially reliable stress features, capturing involuntary physiological changes linked to laryngeal tension and vocal tract resonance, which are less subject to individual variability. Local shimmer, an energy-based feature, serves as a secondary stress marker, reflecting stress-induced reductions in amplitude variability. In contrast, frequency-based features such as jitter and pitch were more affected by sex and individual differences, limiting their reliability for universal stress detection.
Influence of baseline cortisol and sex on speech changes
Baseline speech features provide a crucial reference for evaluating stress-induced vocal changes. Notably, baseline cortisol levels significantly influenced changes in MFCC3 (β = −0.606, p = 0.014) and local shimmer (β = −0.201, p = 0.003), suggesting that individuals with higher baseline cortisol exhibited greater reductions in these features under stress, a pattern previously reported (Carrillo-González et al., Reference Carrillo-González, Atará-Piraquive, Camargo-Mendoza, Hernández-Contreras and Cantor-Cutiva2025). This may reflect reduced vocal flexibility under acute stress, consistent with evidence that elevated baseline cortisol is linked to blunted HPA axis reactivity due to possible desensitisation.
Sex differences also influenced vocal response to stress. Males exhibited lower changes in jitter, shimmer, and pitch-related metrics compared to females, likely due to sex-related differences in neuromuscular control and vocal fold mass. Females generally exhibit more pronounced changes (König et al., Reference König, Riviere, Linz, Elbaum, Fabre and Robert2021), possibly due to more dynamic neuromuscular control over vocal functions and smaller vocal folds, which may lead to greater laryngeal tension and pitch variability. In contrast, males’ thicker vocal folds (Zhang, Reference Zhang2021) may provide greater vocal stability under stress.
State of anger and speech-based stress analysis
Anger affects speech in ways that overlap with, but also differ from, stress-induced vocal changes. In this study, state anger increased significantly in the cold condition but not in the warm condition, suggesting a context-dependent emotional response. Anger was associated with a spectral shift toward lower frequencies, potentially as a compensatory mechanism to enhance vocal dominance. The Hammarberg Index showed a significant negative correlation with STAXI scores in the cold condition, with males exhibiting more pronounced changes than females. These results highlight the need to differentiate anger from stress in speech analysis, as anger uniquely influences vocal intensity and modulation, potentially confounding models designed to detect stress.
Spectral features
Spectral features, which describe how vocal energy is distributed across frequency bands, appear to be the most stable and effective indicators of stress. Alpha ratio and MFCC3 emerged as the most reliable stress biomarkers. Alpha ratio reflects reduced spectral tilt under stress, indicating increased laryngeal tension and reduced high-frequency harmonics, while MFCC3 captures mid-frequency spectral shifts related to stress-induced muscle tension and breath control (Zhang, Reference Zhang2024). These features are particularly robust due to their resistance to voluntary modulation and individual differences, making them suitable for objective stress detection.
A negative correlation between the Hammarberg index and both cortisol levels and STAXI scores in the cold condition suggests stress is associated with a spectral shift toward lower frequencies. This supports prior findings indicating that stress reduces higher frequency energy, possibly due to increased laryngeal tension or altered vocal fold vibration (Bachorowski & Owren, Reference Bachorowski and Owren1995; Laukkanen et al., Reference Laukkanen, Ilomäki, Leppänen and Vilkman2008). The observed sex effects align with research showing males and females exhibit distinct phonatory adjustments under stress, likely due to laryngeal anatomy and hormone-mediated vocal control (Bouhuys et al., Reference Bouhuys, Bloem and Groothuis1995; Fitch & Giedd, Reference Fitch and Giedd1999). Although anger has been linked to increased vocal intensity and pitch modulation (Scherer, Reference Scherer1986; Banse & Scherer, Reference Banse and Scherer1996), our findings indicate that in stress conditions, anger may coincide with a downward spectral energy shift, possibly reflecting a compensatory mechanism or an adaptive response aimed at enhancing vocal dominance (Briefer, Reference Briefer2012).
Energy features: amplitude stability as a robust stress marker
Energy features, which capture variations in speech loudness, amplitude stability, and harmonic content, also provide useful stress indicators. Local shimmer showed a significant decrease under stress (β = −0.201, p = 0.003), indicating that increased laryngeal stiffness stabilises vocal fold vibrations, reducing amplitude variability (Bodaghi et al., Reference Bodaghi, Xue, Thomson and Zheng2025; Carrillo-González et al., Reference Carrillo-González, Atará-Piraquive, Camargo-Mendoza, Hernández-Contreras and Cantor-Cutiva2025; Giddens et al., Reference Giddens, Barron, Byrd-Craven, Clark and Winter2013; Zhang, Reference Zhang2024). The observed increase of mean loudness under stress exposure in part contrasts with previous findings. Dietrich and Verdolini Abbott (Reference Dietrich and Verdolini Abbott2012) reported lower vocal intensity during stress, while Huttunen et al. (Reference Huttunen, Keränen, Väyrynen, Pääkkönen and Leino2011) and Mattei et al. (Reference Mattei, Legou, Cardeau, Le Goff, Lagier and Giovanni2019) found that cognitive load and vocal constraints increased both fundamental frequency and vocal intensity. This discrepancy may arise from the differing study setups. Notably, Dietrich and Verdolini Abbott (Reference Dietrich and Verdolini Abbott2012) only included female participants in their study.
Frequency features: less reliable due to sex and individual variability
Frequency-based measures such as pitch variability and jitter are commonly associated with stress but were found to be less reliable due to significant sex effects and speaker variability. Among them, rap jitter showed some association with stress, but results were inconsistent across individuals. Additionally, pitch mean and pitch min/max were not significantly correlated with cortisol, suggesting these measures are more influenced by voluntary vocal modulation than physiological stress. Our findings align with previous research demonstrating that jitter and other frequency-based measures are influenced by sex (Brockmann et al., Reference Brockmann, Drinnan, Storck and Carding2011). The same study suggests voice sound pressure level (SPL) as a key driver of jitter and shimmer, indicating that the observed sex differences in our study may be partially attributable to variations in voice SPL rather than intrinsic sex effects alone. Additionally, we observed that F2 mean frequency significantly decreased after warm exposure, suggesting a more relaxed vocal tone under low stress. Conversely, F2 frequency SD increased after cold exposure, reflecting greater vocal variability under stress (Protopapas & Lieberman, Reference Protopapas and Lieberman1997). These results support the idea that stress-related physiological changes disrupt vocal stability.
Linear mixed model
Results from our LMM suggest that baseline cortisol and sex are important predictors of changes in specific speech features, with males showing consistently lower changes across jitter, shimmer, and pitch-related metrics, and baseline cortisol levels influencing shimmer and MFCCs. Additionally, changes in cortisol were associated with reduced alpha ratio values, highlighting the physiological impact of cortisol on speech characteristics.
Machine learning results
Classification performance was modest (max AUC = 0.55), suggesting limited discriminability between SECPT groups based on read speech. This likely reflects low expressive variability in structured speech and subtle group-level effects. In contrast, regression models, especially SVMs, performed better. Cortisol MAEs between 3.8–6.0 nmol/L represented ∼ 24–38% of the observed range, STAXI MAEs between 1.3–2.2 points, accounting for ∼ 16–28% of the observed score range. While group-level classification is challenging, acoustic features moderately captured individual differences in stress responses. The most informative features included spectral and temporal voice characteristics such as MFCCs, shimmer, F1 bandwidth and frequency, duration, and alpha ratio. These features reflect articulatory dynamics, phonatory control, and resonance changes, mechanisms plausibly influenced by physiological and emotional stress (Schewski et al., Reference Schewski, Doss, Beldi, Keller and Biassoni2025). Their consistent relevance across tasks highlights the potential of acoustic voice metrics as individual-level stress indicators, even in constrained speech contexts.
Limitations
There are limitations to our study. Recording equipment varied, which may limit comparability of acoustic features across devices. However, it has been demonstrated that different technical devices and recording circumstances have negligible effect on acoustic measures of voice (Awan et al., Reference Awan, Bahr, Watts, Boyer and Budinsky2024; van der Woerd et al., Reference van der Woerd, Wu, Parsa, Doyle and Fung2020). Furthermore, we did not account for participants’ circadian rhythm when assessing cortisol levels, which can influence the results (Stalder et al., Reference Stalder, Kirschbaum, Kudielka, Adam, Pruessner, Wüst, Dockray, Smyth, Evans, Hellhammer, Miller, Wetherell, Lupien and Clow2016). However, all participants were examined in the afternoon, minimising potential confounding effects associated with the cortisol awakening response. Furthermore the time interval between participants’ arrival and the onset of the stressor was not standardised but varied across individuals in the current study, which may have affected baseline cortisol levels at t0. Future research should aim to control and minimise the latency between arrival and the beginning of the experimental procedure to ensure greater comparability across participants. Although the SECPT is a well-established stress induction paradigm, it may not capture the full complexity of stress responses in real-world situations. Lastly, while the 20-minute post-SECPT sample aligns with standard protocols for capturing peak cortisol response, it does not allow us to determine whether speech changes reflect peak stress or early recovery. Future work should include additional time points to better capture the temporal relationship between speech and physiological stress markers.
Conclusion
Our results converge with prior research identifying increased voice pitch and altered spectral properties as markers of stress (Paulmann et al., Reference Paulmann, Furnes, Bøkenes and Cozzolino2016; Kappen et al., Reference Kappen, van der Donckt, Vanhollebeke, Allaert, Degraeve, Madhu, Van Hoecke and Vanderhasselt2022). Our study extends this work by simultaneously accounting for both physiological and psychological stress markers by exploring a broader range of acoustic features. In doing so, we provide a more comprehensive picture of how stress modulates vocal characteristics. Differences between the cold and warm water groups confirm that stress-induced vocal changes are not only measurable but also systematically linked to physiological stress. Speech analysis thus holds promise as a non-invasive, real-time tool for stress assessment, offering advantages over self-report measures or invasive lab-based methods. Features such as jitter, shimmer, and spectral shifts showed sensitivity to stress-related fluctuations, making them ideal for continuous monitoring. Importantly, speech reflects dynamic physiological states and could aid in detecting both acute and chronic stress, conditions closely tied to mental health. Voice-based markers may also support early detection and management of stress-related disorders like anxiety and depression (Tafet & Nemeroff, Reference Tafet and Nemeroff2016; Daviu et al., Reference Daviu, Bruchas, Moghaddam, Sandi and Beyeler2019) (Gaikwad & Venkatesan, Reference Gaikwad, Venkatesan, Jha, Tripathi, Natarajan and Sharma2024). By correlating speech patterns with cortisol, clinicians gain real-time insight into stress, enabling earlier and more objective interventions. Not all acoustic features are equally reliable for stress detection, underscoring the importance of selecting robust and objective acoustic features for stress detection.
Future analyses should consider replicating existing findings in larger, more diverse populations and across different stress-inducing contexts. Importantly, since our current approach relies on pre/post comparisons relative to a defined baseline, future work should explore how speech-based stress detection can be adapted for real-world settings where baseline definitions are less straightforward. This could involve passive data collection to estimate individual baselines or models that operate without requiring baseline data. Additionally, integrating speech analysis with other physiological measures and exploring its use in clinical practice could further refine its potential for real-time stress monitoring.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/neu.2025.10037.
Acknowledgements
We would like to thank all participants of this study.
Author contribution
FM contributed to conceptualisation and study design, data analysis and interpretation, manuscript drafting, critical revision, and project administration; HL contributed to conceptualisation and study design, data analysis and interpretation, manuscript drafting, and critical revision; JT contributed to data analysis and interpretation, manuscript drafting, critical revision, supervision, and project administration; SP contributed to manuscript drafting, critical revision; AK contributed to manuscript drafting, critical revision, and supervision; NS contributed to data acquisition, manuscript revision, and project administration; AR contributed to conceptualisation, and funding acquisition; MMP contributed to conceptualisation, data acquisition, analysis and interpretation, manuscript drafting and revision, funding acquisition, supervision, and project administration; MS contributed to conceptualisation, data acquisition, analysis and interpretation, manuscript drafting and revision, funding acquisition, supervision, and project administration. All authors approved the final version of the manuscript.
Financial support
This work was supported by the Collaborative Research Center (TRR379) of the German Research Foundation (DFG) – Project number: 512007073; and the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) LORA-Risk, Grant 519365575 to Michael M. Plichta and Andreas Reif.
Competing interests
FM, HL, JT and AK are employees of the speech biomarker company ki:elements GmbH, Saarbrücken, Germany. The remaining authors have nothing to disclose.
Ethical statement
This experiment is part of the larger study, ‘Identifying mediators of stress-aggression and experimental manipulation by tDCS’, approved by the Ethics Committee of the Goethe University Frankfurt Medical Faculty (Ref: 20-1035). The study was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent and received financial compensation of €10 per hour for their participation.
Open access
