Background
Anxiety disorders are the most prevalent mental health disorders in adolescents and are associated with adverse psychosocial outcomes (Rockhill et al., Reference Rockhill, Kodish, DiBattisto, Macias, Varley and Ryan2010). They often manifest early, with a mean age of onset of 11 years (Kessler et al., Reference Kessler, Berglund, Demler, Jin, Merikangas and Walters2005), yet they remain largely undertreated. When left untreated, anxiety disorders often become chronic and predict other mental health conditions in adulthood (Essau, Conradt, & Petermann, Reference Essau, Conradt and Petermann2002). One critical factor in the underrecognition of anxiety disorders in young people is the perception that anxiety symptoms are intrinsic personality traits rather than treatable conditions (Wang, Berglund, Olfson, & Kessler, Reference Wang, Berglund, Olfson and Kessler2004). Additionally, many adolescents are reluctant to discuss their anxiety symptoms and seek help due to stigma concerns (Gulliver, Griffiths, & Christensen, Reference Gulliver, Griffiths and Christensen2010). This is unfortunate, as effective, non-stigmatizing treatments for anxiety disorders are available (Siegel & Dickstein, Reference Siegel and Dickstein2011). A low-intensity, low-cost screening method for detecting anxiety disorders that does not require young people to verbalize their symptoms could be a first step in the early detection of anxiety disorders, helping those who need professional treatment and thereby preventing the stagnation of their social and educational development.
One potential non-confrontational screening method for anxiety disorders is through the automatic extraction of acoustic-prosodic markers from speech. Despite their heterogeneity, anxiety disorders consistently exhibit a strong physical component. Anxiety increases muscle tension and respiratory rate (Horvath & Fenz, Reference Horvath and Fenz1971), altering the mechanics of speech production and, consequently, the acoustic-prosodic properties of speech sounds (Sondhi et al., Reference Sondhi, Khan, Vijay, Salhan and Chouhan2015).
Several studies have identified cross-sectional associations between anxiety and speech characteristics in adults. Anxiety often affects the fundamental frequency (F0) of speech, which is the rate at which the vocal cords vibrate (open and close) across the glottis during phonation, perceived as pitch. Most studies have found that mean F0 increases with anxiety (Hagenaars & van Minnen, Reference Hagenaars and van Minnen2005; Laukka et al., Reference Laukka, Linnman, Åhs, Pissiota, Frans, Faria and Furmark2008; Teferra et al., Reference Teferra, Borwein, DeSouza, Simpson, Rheault and Rose2022; Weeks et al., Reference Weeks, Lee, Reilly, Howell, France, Kowalsky and Bush2012), although others report no change (Goberman, Hughes, & Haydock, Reference Goberman, Hughes and Haydock2011; Wortwein, Morency, & Scherer, Reference Wortwein, Morency and Scherer2015). Anxiety-induced muscle tension can also alter formant frequencies (F1, F2, F3), which define vowel sounds by reflecting the concentration of acoustic energy at specific frequencies of the vocal tract (Özseven, Düğenci, Doruk, & Kahraman, Reference Özseven, Düğenci, Doruk and Kahraman2018; Wortwein et al., Reference Wortwein, Morency and Scherer2015). Temporal aspects of speech, such as speech rate and pause frequency, may also be influenced by anxiety. Some studies indicate that anxious individuals speak faster and pause less (Murray & Arnott, Reference Murray and Arnott1993; Siegman & Boyle, Reference Siegman and Boyle1993), while others suggest the opposite, with higher anxiety associated with slower speech and more frequent pauses (Goberman et al., Reference Goberman, Hughes and Haydock2011; Hagenaars & van Minnen, Reference Hagenaars and van Minnen2005; Laukka et al., Reference Laukka, Linnman, Åhs, Pissiota, Frans, Faria and Furmark2008; Teferra et al., Reference Teferra, Borwein, DeSouza, Simpson, Rheault and Rose2022;). Anxiety has also been associated with voice quality as assessed by features such as jitter (cycle-to-cycle F0 variation), shimmer (cycle-to-cycle loudness variation), and Mel-Frequency Cepstral Coefficients (MFCC) (spectral properties of speech sounds), but findings vary across studies (Jones, Anagnostou, & Verhoeven, Reference Jones, Anagnostou and Verhoeven2011; Özseven et al., Reference Özseven, Düğenci, Doruk and Kahraman2018; Teferra et al., Reference Teferra, Borwein, DeSouza, Simpson, Rheault and Rose2022). All speech deviations in anxious individuals may be exacerbated under stress, such as during public speaking situations.
So far, no study has investigated this relationship among adolescents, a critical period for anxiety onset. Adolescent speech differs from that of adults due to rapid developmental changes affecting their voice, including the widening of the larynx and elongation of the vocal tract (Zamponi, Mazzilli, Mazzilli, & Fantini, Reference Zamponi, Mazzilli, Mazzilli and Fantini2021). Compared to adults, they exhibit higher pitch and formant frequencies, longer durations of individual phonetic segments, and greater variability in temporal and spectral features (Lee, Potamianos, & Narayanan, Reference Lee, Potamianos and Narayanan1999). Thus, anxiety may manifest differently in adolescent speech, potentially necessitating specific acoustic markers for accurate detection.
Furthermore, voice changes in adolescents follow different developmental trajectories in males and females, with males experiencing more pronounced transformations (Zamponi et al., Reference Zamponi, Mazzilli, Mazzilli and Fantini2021). Male adolescents experience substantial decreases in F0 and formant frequencies, resulting in lower-pitched, more resonant voices compared to females (Markova et al., Reference Markova, Richer, Pangelinan, Schwartz, Leonard, Perron and Paus2016).
The above-described findings show that, although several studies have assessed associations between acoustic-prosodic markers and anxiety in adults, the direction of effects has remained inconclusive, with most studies relying on cross-sectional analyses. Assessing the predictive potential of speech quantification in relation to the risk of developing anxiety in the future is needed for understanding its value in screening and early detection.
In the present study, we evaluated cross-sectional and longitudinal associations between anxiety and acoustic-prosodic markers, using data from a longitudinal cohort of Dutch adolescents collected during a standardized social stress test. First, we conducted an exploratory, sex-specific analysis of the associations between speech patterns and anxiety symptoms in adolescents around age 16. Subsequently, we evaluated the potential of using sex-specific acoustic markers assessed at age 16 to predict the severity of anxiety symptoms and the diagnosis of an anxiety disorder at age 19. Given the nature of the speech elicitation task – a test designed to induce stress in socially evaluative situations – we hypothesize that this task will be particularly effective at eliciting speech markers specific to social anxiety disorder (SAD), thereby yielding greater accuracy in detecting SAD than a composite anxiety diagnosis including various anxiety disorders. Additionally, drawing on previous research (Weeks et al., Reference Weeks, Lee, Reilly, Howell, France, Kowalsky and Bush2012), we hypothesize that SAD will be more easily recognized through acoustic speech markers in male than in female adolescents.
Methods
Participants
This study is based on data from a subsample of 581 participants (51.1% females) selected from the longitudinal population-based TRacking Adolescents’ Individual Lives Survey (TRAILS) (Oldehinkel et al., Reference Oldehinkel, Rosmalen, Buitelaar, Hoek, Ormel, Raven and Hartman2015), which includes biennial or triennial assessments from age 11 onwards. The total sample comprised 2230 participants, of whom 715 agreed to participate in a series of additional tests, including a digitally recorded speech task (Oldehinkel et al., Reference Oldehinkel, Ormel, Bosch, Bouma, Van Roon, Rosmalen and Riese2011). Of these speech recordings, 581 were of sufficient quality for quantitative analysis (see subsection 2.3 below). The data were collected when participants were aged 15–17 years (mean = 15.7, SD = 0.67) and 18–20 years (mean = 18.5, SD = 0.59) respectively. Inclusion criteria required consent of the child’s primary school to participate in the study. Children were excluded if they had a severe intellectual disability, a significant physical illness or impairment, or if no Dutch-speaking parent or guardian was available to assist with study procedures. For a full description of the inclusion and exclusion criteria, see de Winter et al. (Reference de Winter, Oldehinkel, Veenstra, Brunnekreef, Verhulst and Ormel2005). Written informed consent was obtained from both parents and adolescents prior to enrollment. The study was approved by the Dutch Central Committee on Research Involving Human Subjects (CCMO) and adhered to the principles of the Declaration of Helsinki.
Speech elicitation task
At baseline, participants completed the speech task from the Groningen version of the ‘Trier Social Stress Test’ (GTSST), a protocol originally designed to create a moderate level of performance-related psychological stress in a lab environment using various tasks (Kirschbaum, Pirke, & Hellhammer, Reference Kirschbaum, Pirke and Hellhammer2008). For the GTSST, participants were instructed in loco to prepare and deliver a monologue, in which they had to speak continuously for 6 minutes about themselves and their lives. Participants delivered the monologue in front of both a video camera and a research assistant. Prior to the task, they were informed that their recorded performance would later be evaluated and ranked by a panel of peers based on speech content, use of voice, and posture. During the task, the research assistant observed the participant with a critical disposition. The speech task took place in test rooms located near participants’ residences (e.g., hotels, community centers). All rooms were soundproof with blinded windows. Built-in microphones integrated with video cameras were consistently used for recording. The experimental setup (i.e., position of the camera in relation to the participant) was similar in all sessions.
Acoustic analyses
Each participant’s audio channel was extracted from the video recording using the FFmpeg library (https://ffmpeg.org/). Audio quality was assessed by a trained researcher who manually inspected spectrograms using Praat (Boersma & Weenink, Reference Boersma and Weenink2013), checking for clipping, background noise, and human intelligibility.
The 581 audio recordings were acoustically analyzed with OpenSMILE (version 2.0), employing the extended Geneva Acoustic Minimalistic Parameter Set (eGeMAPS) (Eyben, Weninger, Gross, & Schuller, Reference Eyben, Weninger, Gross and Schuller2013). This set contains 88 acoustic parameters, sorted into four categories: frequency parameters (28 parameters, e.g., fundamental frequency), energy/amplitude parameters (10 parameters, e.g., loudness), spectral balance parameters (44 parameters, e.g., formants), and temporal parameters (6 parameters, e.g., speech rate). Parameters were derived from 18 low-level descriptors, for which the arithmetic mean and coefficient of variation were calculated. For pitch and loudness, eight additional measures were computed: 20-th, 50-th, and 80-th percentiles, percentile range (20-th to 80-th), and mean and standard deviation of falling and rising slope. These additional measures, part of the eGeMAPS set (Eyben et al., Reference Eyben, Scherer, Schuller, Sundberg, Andre, Busso and Truong2016), capture both distributional (percentiles) and dynamic (slopes) aspects of pitch and loudness.
Anxiety measures
Severity of anxiety symptoms was evaluated at baseline using the Youth Self Report ([YSR]; Achenbach & Rescorla, Reference Achenbach and Rescorla2001) and at follow-up using the Adult Self Report ([ASR]; Achenbach & Rescorla, Reference Achenbach and Rescorla2003). The YSR includes six items and the ASR includes seven items related to anxiety symptoms. For both the YSR and ASR, the anxiety scale scores were calculated by averaging the anxiety-related item scores. The YSR demonstrated moderate internal consistency with a Cronbach’s alpha of 0.65, while the ASR showed good internal consistency with a Cronbach’s alpha of 0.75.
At follow-up, the World Health Organization Composite International Diagnostic Interview (CIDI 3.0) was used to assess the presence of anxiety disorders according to the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV) (Haro et al., Reference Haro, Arbabzadeh-Bouchez, Brugha, de Girolamo, Guyer, Jin and Kessler2006). Diagnoses included lifetime occurrences of adult separation anxiety, agoraphobia, generalized anxiety disorder, panic disorder, childhood separation anxiety, social anxiety, and specific phobia. Individuals diagnosed with an obsessive-compulsive disorder were excluded from this study as the DSM-5 reclassifies this condition outside the category of anxiety disorders (2013) (American Psychiatric Association, 2013). A composite anxiety disorder measure was created based on the presence of any of the seven specified diagnoses, with participants having at least one of these disorders being classified as having an anxiety disorder (hereafter referred to as AD).
Statistical analyses
Data transformation
Boxplots based on the interquartile range (IQR) were used to visually inspect the distributions of acoustic features. To quantify the asymmetry of each distribution, the skewness coefficient was calculated for each feature. Twenty-four acoustic variables with skewed distributions (skewness coefficient above +1 or below −1) were transformed to better fit a normal distribution (Supplementary Table S1). A logarithmic transformation was applied to variables with mild positive skewness (skewness coefficient between +1 and + 3), and a reciprocal transformation (1/x) was applied to those with severe positive skewness (skewness coefficient greater than +3). For negatively skewed variables (skewness coefficient less than −1), square transformation was applied. Variables that remained skewed (skewness coefficient greater than +1 or less than −1) after transformation were excluded, resulting in the removal of five variables, leaving 83 parameters for analysis (for a list, see Supplementary Table S2). All statistical analyses were performed with R (version 4.1.2) (R Core Team, 2020).
Cross-sectional associations with anxiety symptoms
Independent two-tailed bivariate Spearman’s correlations were calculated between each of the 83 speech features and baseline anxiety severity to investigate cross-sectional relationships. The Benjamini-Hochberg (BH) method was employed to adjust for the high false discovery rate (FDR) due to multiple correlation tests. Analyses were performed separately for males and females.
Longitudinal associations with anxiety symptoms
Backward linear regression models were used to test the longitudinal relationship between speech features and anxiety severity at follow-up. To avoid multicollinearity, pairs of highly intercorrelated acoustic features (i.e., Spearman’s rho > .85) were identified (Elith et al., Reference Elith, Graham, Anderson, Dudík, Ferrier, Guisan and Zimmermann2006), and the feature with the highest average correlation to other features was excluded (Supplementary Table S3). The backward elimination process, based on the Akaike information criterion (AIC), systematically removed one predictor at a time until further removals lowered the AIC. Standardized regression coefficients (Beta) were used to rank predictors by their relative importance.
To assess the predictive effect of speech features on follow-up anxiety beyond baseline anxiety, a linear regression model was first constructed using baseline anxiety as the sole predictor. In a second model, acoustic features were added to baseline anxiety. The fits of the two models were compared by means of F-statistics. Analyses were performed separately for male and female participants.
Longitudinal classification of anxiety disorders
Binary Random Forest (RF) classifiers (Breiman, Reference Breiman2001) were used to distinguish individuals with a lifetime anxiety disorder (AD) and a social anxiety disorder (SAD) from healthy controls based on speech features. To create balanced groups, the dataset was down sampled to include 246 participants for AD (123 with AD, 123 without) and 108 for SAD (54 with SAD, 54 without), matched for sex, age, and education. All speech features were standardized to have a mean of 0 and a standard deviation of 1 before being used in the model. All models were trained using 10-fold cross-validation, where in each iteration, the model was trained on data from 9 folds and tested on the remaining fold. This process was repeated 10 times, ensuring that each fold served as a test set once. Classification was performed using a subset of features selected through Recursive Feature Elimination with Cross-Validation (RFECV). RFECV iteratively removed the least informative features one at a time based on accuracy, using 10-fold cross-validation. Gini importance scores were computed to compare the relative ranking of acoustic features in the classifiers. Model performance was evaluated using accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC-ROC). All analyses were conducted separately for males and females.
Results
Descriptive statistics
Table 1 presents demographic and clinical information of the participants, stratified by sex. Female participants exhibited higher levels of anxiety at both baseline and follow-up compared to males (Table 1). A total of 123 participants met criteria for at least one anxiety disorder (AD)(Supplementary Table S4). This group included 54 individuals diagnosed with social anxiety disorder (SAD). Of those with SAD, 21 also met criteria for at least one additional AD. Educational level did not differ between males and females (Table 1), nor between anxious and non-anxious participants (Supplementary Table S4).
Table 1. Demographic and clinical characteristics of participants stratified by sex
Note: Reported values are means ± SD or n (%).
Abbreviations: ASR, adult self report; N, sample size; y, years; YSR, youth self report.
a Percentages exceed 100% because some individuals were diagnosed with multiple anxiety disorders.
Cross-sectional associations with anxiety symptoms
Significant associations were identified between baseline anxiety and four acoustic-prosodic features in males and one in females (Supplementary Figure S1). However, none of these correlations remained statistically significant after false discovery rate correction for multiple comparisons (based on 83 tests). In males, three of the associated features pertained to pitch and formants, while one was related to voice quality. Higher anxiety severity was negatively correlated with the mean pitch falling slope (rho = −0.166, p = .006), pitch range (20th to 80th percentile) (rho = −0.189, p = .002), and mean frequency of the first formant (F1) (rho = −0.127, p = .036), indicating that increased anxiety is associated with a slower pitch falling rate, a narrower pitch range, and a lower F1 frequency. Conversely, anxiety was positively correlated with the coefficient of variation of MFCC1 (rho = 0.132, p = .029). In females, a negative association was observed between the standard deviation of the pitch falling slope and anxiety (rho = −0.119, p = .043), indicating that a lower variation in the rate at which pitch falls is linked to higher anxiety severity (Supplementary Figure S1).
Longitudinal associations with anxiety symptoms
After removing highly intercorrelated features (i.e., Spearman’s rho > .85), 55 speech parameters were analyzed using backward linear regression models to predict anxiety levels at 3-year follow up, separately for males and females. In male participants, the optimal model identified nine features, accounting for 24.6% of the variance (Table 2). Adding speech features resulted in a 5.4% increase in explained variance compared to a model with only baseline anxiety. This increase was statistically significant (F(8, 238) = 2.567, p = .002). For females, the best-fitting regression model included 23 predictors which explained 33.2% of the variance in future anxiety (Table 2). Adding speech features to a model that already accounted for baseline anxiety resulted in a 10.9% increase in explained variance. This improvement was statistically significant (F(22, 253) = 3.026, p < .001). The top three predictors of future anxiety severity for males and females are presented in Table 3 (for a full list, see Supplementary Table S5).
Table 2. Linear regression results predicting anxiety severity at 3-year follow-up using acoustic-prosodic features, baseline anxiety, and a combined model including both
Note: Best performing model (highest R 2, lowest RSE) in bold.
Abbreviations: N, number; RSE, residual standard error.
Table 3. Top three variables predicting anxiety severity at 3-year follow-up in male and female adolescents, ranked by absolute value of standardized coefficients (highest to lowest)
Abbreviations: β, standardized coefficient; b, unstandardized coefficient; SE, standard error.
Longitudinal classification of anxiety disorders
The demographic characteristics of the matched subset of 246 participants, used for the random forest classifiers, are presented in Supplementary Table S4. The RF model achieved the highest performance in distinguishing individuals with a lifetime diagnosis of SAD at age 19, with an AUC-ROC of 85% for males (N = 44) and 74% for females (N = 64) (Table 4). The accuracy was 74% for males and 68% for females. The top 10 discriminative features for SAD were primarily related to pitch and voice quality in males, whereas in females, they were associated with voice quality, pitch, and pauses (Figure 1). Additionally, the RF model achieved an AUC–ROC of 73% for males (N = 84) and 67% for females (N = 162) when classifying adolescents with a lifetime diagnosis of any anxiety disorder (AD) at age 19 (Table 4). The accuracy was 68% for males and 61% for females. The complete list of features selected for SAD and AD classification can be found in the supplementary materials (Supplementary Table S6).
Table 4. Classification performance of random forest models predicting anxiety (AD) and social anxiety disorder (SAD) diagnoses at 3-year follow-up in male and female adolescents
Note: Best performing model in bold.
Abbreviations: N, number; AUC–ROC, area under the curve–receiver operating characteristic.
Figure 1. Top 10 acoustic-prosodic features ranked by Gini importance for the longitudinal classification of Social Anxiety Disorder (SAD) in males (a) and females (b). Feature importance was computed as the total decrease in Gini impurity per feature across all trees in the random forest classifier.
Discussion
This is the first study to assess cross-sectional and longitudinal associations between acoustic-prosodic markers and anxiety in a large sample of male and female adolescents representative of the general (Dutch) population. The model accurately identified SAD three years in advance in males, achieving an AUC–ROC of 85%, and demonstrated moderate accuracy in females, with an AUC–ROC of 74%. Speech features improved the clinical prediction of anxiety severity at three years, with distinct acoustic predictors identified for males and females in the linear regression models.
Cross-sectional associations with anxiety symptoms
Several acoustic features were cross-sectionally correlated with anxiety severity; however, none remained statistically significant after correction for multiple testing and they should thus be interpreted with caution. In both sexes, anxiety severity was negatively associated with the falling slope of F0 (pitch). Pitch contours, characterized by rising and falling slopes, contribute to intonation and convey information about linguistic meaning, syntactic structures, and speaker emotions (Laukka, Reference Laukka, Zeigler-Hill and Shackelford2017; Vaissiére, Reference Vaissiére, Pisoni and Remez2008). Females with higher anxiety exhibited a falling pitch trajectory with reduced variability, while anxious males showed a slower pitch falling rate. Both results suggest reduced pitch dynamism, consistent with emotion regulation theories proposing that anxious individuals suppress their emotional expressiveness, particularly in socially evaluative situations like the Trier Social Stress Test used in this study (Gross, Reference Gross2002).
In males, higher anxiety was associated with a narrower pitch range, indicating a reduced span of vocal frequencies in their speech. This aligns with previous research by Drioli, Tisato, Cosi, and Tesser (Reference Drioli, Tisato, Cosi and Tesser2003) which demonstrated a lower pitch range in a male university student during automatic fear detection compared to other emotions. No such effect emerged in females, potentially due to their naturally wider pitch range (195–275 Hz in females vs. 140–214 Hz in males; Wilson, Reference Wilson1987), which may make anxiety-related increases harder to detect.
The frequency of the first formant (F1) was negatively correlated with anxiety in males. F1 reflects the physical characteristics of the vocal tract, and is negatively correlated with tongue height during speech (Lee, Shaiman, & Weismer, Reference Lee, Shaiman and Weismer2016). Anxiety-induced tension in the supraglottal articulators can alter their movement and positions, influencing F1 values. This partly aligns with previous research showing reduced F1 in anxious states, although without sex-specific effects (Özseven et al., Reference Özseven, Düğenci, Doruk and Kahraman2018). Another study observed a correlation between anxiety and F1 during a social stress test in adult males but not in adult females (Teferra et al., Reference Teferra, Borwein, DeSouza, Simpson, Rheault and Rose2022), just like we did, but in the opposite direction. This discrepancy may reflect age differences between the study samples (mean ages of approximately 16 and 36). During adolescence, anatomical changes in the male vocal tract lower F1 frequency, with maturation continuing until around age 19 (Lee et al., Reference Lee, Potamianos and Narayanan1999). These changes may modulate the anxiety–F1 relationship in adolescents. Lastly, MFCC1 variability positively correlated with anxiety in males, suggesting more fluctuations in energy distribution across frequencies, likely due to articulatory instability from supraglottal muscle tension. Given that the task was self-referential and likely to elicit social-evaluative concerns – a central feature of social anxiety – future studies should investigate associations between state social anxiety during the TSST and acoustic markers, to better disentangle these effects from those of trait anxiety.
Longitudinal associations with anxiety symptoms
Our longitudinal analyses demonstrated that acoustic features predicted anxiety severity at age 19, after accounting for baseline anxiety at age 16. The additional variance explained by acoustic features (5.4% for males, 10.9% for females) demonstrates that speech markers provide incremental predictive value beyond baseline anxiety levels. This suggests that these acoustic features capture information not redundant with traditional self-reported measures of anxiety, enhancing predictive precision.
Acoustic features predictive of future anxiety severity varied by sex, confirming our initial hypothesis that accurate prediction of anxiety disorders may benefit from developing sex-specific prediction models. Loudness emerged as the strongest acoustic predictor of future anxiety in male adolescents. Males with higher anxiety at age 19 exhibited more loudness peaks in their speech and a slower rate of loudness decline at age 16. Changes in loudness are one of the main characteristics contributing to the perception of instability in the voice of male adolescents (Boltez̆ar, Burger, & Z̆argi, Reference Boltez̆ar, Burger and Z̆argi1997), often attributed to the growth of their respiratory system and the necessary adaptation of nervous control over the speech apparatus. Increased loudness has also been linked to heightened emotional arousal (Siegman & Boyle, Reference Siegman and Boyle1993). Together, these results suggest that a stressful situation may exacerbate loudness instabilities in individuals vulnerable to developing anxiety.
Females with higher levels of anxiety at age 19 spoke faster with more alterations in vocal quality. This aligns with studies linking anxiety to a faster speech rate (Murray & Arnott, Reference Murray and Arnott1993; Siegman & Boyle, Reference Siegman and Boyle1993) and highlights its potential as an early marker of anxiety in female adolescents. Women generally experience increased arousal and quicker responses to stress compared to men (Bangasser, Eck, Telenson, & Salvatore, Reference Bangasser, Eck, Telenson and Salvatore2018), which may help explain their faster speech rate in such situations. In females predisposed to anxiety, this response may become maladaptive and predict greater future anxiety severity (Dieleman et al., Reference Dieleman, Huizink, Tulen, Utens and Tiemeier2016; Nelemans et al., Reference Nelemans, Hale, Branje, van Lier, Koot and Meeus2017). Overall, these findings suggest that speech patterns not only reflect current anxiety but may also serve as early markers of increased vulnerability to developing anxiety over time.
Longitudinal classification of anxiety disorders
The RF model trained on selected feature subsets achieved the highest AUC–ROC for the longitudinal classification of individuals with a lifetime diagnosis of SAD: 85% for males and 74% for females. Notably, these findings were derived from a non-clinical, non-help-seeking population, reinforcing the potential of speech markers for the early detection of SAD in the general population. Despite females being more likely to receive a diagnosis of SAD and reporting more severe symptoms than males, the speech-based automatic classification model had greater accuracy in males. The most informative feature for classifying SAD in males was the falling slope of pitch, which we also found to correlate cross-sectionally with anxiety severity. These findings align with prior research showing that pitch during social threat tasks can serve as a physiological marker of SAD in males, with effects that do not always generalize or are weaker in females (Weeks et al., Reference Weeks, Lee, Reilly, Howell, France, Kowalsky and Bush2012). From a developmental perspective, Scharfstein et al. (Reference Scharfstein, Beidel, Sims and Rendon Finnell2011) found that children with SAD had higher pitch during a role play assessment than children with Asperger’s disorder, though neither differed significantly from typically developing peers. Our results support and extend these findings by showing that, during adolescence, the association between pitch and SAD relative to healthy groups becomes more pronounced in males, with a more modest effect observed in females, as indicated by the selected features in the RF model. Future studies could examine whether pubertal voice maturation helps explain why pitch-based features are more predictive of SAD in boys than in girls. In males, puberty is accompanied by a substantial drop in pitch, and delays in this process have been linked to increased peer-related stress and social withdrawal (Dwyer, Reference Dwyer2020), which may increase vulnerability to social anxiety. We recommend that future research include measures of vocal maturation to test this hypothesis.
As hypothesized, the SAD classification model outperformed the overall anxiety model (AUC-ROC of 72.4% for males and 63.3% for females). This finding may be attributed to the heterogeneity of anxiety disorders, which involve diverse symptom profiles and triggers (Drzewiecki & Fox, Reference Drzewiecki and Fox2024). While social anxiety is primarily triggered by fear of social evaluations, other anxiety disorders may be influenced by different stimuli, such as specific situations or sensory challenges, which were not elicited in this study. The superior performance of the SAD model may be due to the fact that the Trier Social Stress Test is designed to elicit stress responses that may particularly trigger individuals with SAD, in turn eliciting speech responses that directly tap into this disorder. Consistent with this, prior research has shown that the association between pitch and SAD in a socially evaluative task was not attributable to general arousal or symptoms of generalized anxiety or panic disorder but was specific to SAD (Weeks et al., Reference Weeks, Lee, Reilly, Howell, France, Kowalsky and Bush2012). These findings underscore the critical role of task design in detecting speech markers for anxiety disorders.
Limitations
This study has some limitations that warrant consideration. First, dividing the sample by sex reduced the statistical power, particularly in the analysis of anxiety subtypes, highlighting the need for replication in larger samples to validate the findings. Second, after normalizing the correlation results across 83 tests, none of the correlation analyses reached statistical significance. This outcome was anticipated given the high number of features examined, which increased the likelihood of non-significant results after correction for multiple testing. Future studies may benefit from feature reduction techniques to mitigate this challenge. A third limitation is that, although groups did not differ in educational level – a proxy for cognitive ability (Huh et al., Reference Huh, Arpawong, Gruenewald, Fisher, Prescott, Manly and Gatz2024) – we did not assess whether cognitive functioning moderates the relationship between speech features and anxiety, which could be explored in future research. Similarly, future research may account for comorbid conditions such as Autism Spectrum Disorder (ASD) or depression, which may influence both speech production and anxiety levels.
Last, we used lifetime anxiety disorders that could have started before the assessment of speech, as power restrictions did not allow focusing on onsets after the speech assessment only. However, we did adjust for pre-assessment anxiety in the analyses with symptom severity as outcome measure and these analyses still indicated a prospective effect. Future studies may prospectively follow adolescents without a baseline diagnosis and collect repeated speech samples alongside clinical assessments to assess whether acoustic features can predict the onset of anxiety disorders.
Conclusion
The current study demonstrates that acoustic speech markers elicited in a socially evaluative context are most predictive for the early detection of SAD in male adolescents, with modest accuracy in females and in other anxiety disorders. Speech analyses may thus aid in recognizing vulnerable adolescents at risk of developing anxiety disorders, facilitating early intervention.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291725102274.
Acknowledgements
We are grateful to everyone who participated in this research or worked on this project to make it possible.
Author contribution
S. Ciampelli: Conceptualization, Methodology, Formal analysis, Writing – original draft. J.N. de Boer: Conceptualization, Methodology, Writing – review & editing, Supervision. S. Koops: Conceptualization, Methodology, Writing – review & editing. A.E. Voppel: Conceptualization, Methodology, Writing – review & editing. A.J. Oldehinkel: Conceptualization, Methodology, Writing – review & editing, Supervision, Funding acquisition. I.E.C. Sommer: Conceptualization, Methodology, Writing – review & editing, Supervision. All authors contributed to the manuscript by providing critical revisions, participating sufficiently in the work, and approving the final version for submission.
Funding statement
This research is part of the TRacking Adolescents’ Individual Lives Survey (TRAILS). Participating centers of TRAILS include the University Medical Center and University of Groningen, the University of Utrecht, and the Parnassia Psychiatric Institute, all in the Netherlands. TRAILS has been financially supported by various grants from the Netherlands Organization for Scientific Research NWO (Medical Research Council program grant GB-MW 940–38-011; ZonMW Brainpower grant 100–001-004; ZonMw Risk Behavior and Dependence grant 60–60600–97-118; ZonMw Culture and Health grant 261–98-710; Social Sciences Council medium-sized investment grants GB-MaGW 480–01-006 and GB-MaGW 480–07-001; Social Sciences Council project grants GB-MaGW 452–04-314 and GB-MaGW 452–06-004; ZonMw Longitudinal Cohort Research on Early Detection and Treatment in Mental Health Care grant 636340002; NWO large-sized investment grant 175.010.2003.005; NWO Longitudinal Survey and Panel Funding 481–08-013 and 481–11-001; NWO Vici 016.130.002, 453–16-007/2735, and Vi.C.191.021; NWO Gravitation 024.001.003), the Dutch Ministry of Justice (WODC), the European Science Foundation (EuroSTRESS project FP-006), the European Research Council (ERC-2017-STG-757364 and ERC-CoG-2015-681466), Biobanking and Biomolecular Resources Research Infrastructure BBMRI-NL (CP 32), the Gratama foundation, the Jan Dekker foundation, the participating universities, and Accare Centre for Child and Adolescent Psychiatry.
Competing interests
All authors declare that they have no conflicts of interest.
Ethical standard
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
