This study received approval from the ethics committee of the Fourth People’s Hospital of Chongqing (approval 2025; ethical review 73). Informed consent was obtained from all participants and/or their parents or legal guardians, in line with the Declaration of Helsinki. Participants with SI were further evaluated and intervened accordingly.

A total of 54 university students aged 18 to 25 years were recruited via an online survey platform to rate candidate sentences. These participants did not take part in stage 2. A pool of 50 candidate sentences was created to cover three domains: (1) positive-valence statements, (2) negative-valence expressions, and (3) neutral factual statements. The positive-valence statements were imbued with Chinese traditional virtues and socialist core values to strengthen cultural resonance. To elicit diverse prosodic patterns, sentences were varied across syntactic moods (declarative, interrogative, and exclamatory). Sentence length was restricted to 100 Chinese characters or less, with a reading duration of less than 30 seconds to maintain feasibility and reduce fatigue.

Participants rated all 50 sentences on 9-point Likert scales for pleasure, arousal, and dominance, generating quantitative affective profiles. Two licensed clinical psychologists independently evaluated the sentences for readability, emotional expressiveness, and categorical balance. On the basis of consensus, 12 sentences (4 per valence domain) were selected. Four suicide-related statements adapted from validated instruments [26] were then added, resulting in a final set of 16 standardized prompts (Table S1 in Multimedia Appendix 1).

We conducted a case-control study among university students in southwest China using a sequential screening-and-confirmation process. Data were collected via a WeChat mini program, yielding 747 questionnaire responses and 771 speech recording submissions. Speech recordings were collected on participants’ personal smartphones in familiar, low-stress environments to enhance comfort and ecological validity [27]. Raw audio files were stored in encrypted, access-controlled servers and were used only for acoustic feature extraction [28]. After excluding incomplete entries and screening for audio quality, 715 participants remained eligible for SI screening. SI was initially screened using the Self-Rating Idea of Suicide Scale (SIOSS; threshold ≥12). To reduce misclassification due to concealment, individuals scoring 4 or more on the SIOSS concealment subscale were excluded. Participants who were screened as positive then underwent counselor-administered clinical confirmation using the Columbia-Suicide Severity Rating Scale, requiring an ideation subscale score of 2 or more. This process identified 48 clinically confirmed SI cases. From those who were screened as negative on the SIOSS (<12), we randomly selected 48 controls (1:1) using a computer-generated random sample. Age and sex were compared between groups and were adjusted for in subsequent analyses. The final analytic sample comprised 96 participants. Each participant recorded 16 standardized utterances in quiet settings, yielding 1536 recordings for model development.

All recordings underwent a standardized preprocessing pipeline. Audio was first enhanced using the Model Scope Acoustic Noise Suppression model. To verify reading compliance and exclude off-target recordings, speech was transcribed using the Google Speech-to-Text API (automatic language detection enabled). Transcripts were aligned with target sentences using a dual-weighted fuzzy matching procedure (FuzzyWuzzy library; 60% token set similarity and 40% edit distance similarity); a composite similarity score above 65% indicated a valid match. All matches were manually reviewed, and ambiguous cases were adjudicated by trained researchers. Transcripts were used solely for quality control and compliance checking; no linguistic features were included in model training. Recordings were then converted from MP3 (44.1 kHz; stereo) to WAV (16 kHz; mono; 16 bits) file formats using FFmpeg. After preprocessing, downstream analyses were conducted exclusively on deidentified acoustic feature values rather than raw speech signals.

Acoustic features were extracted using openSMILE[29] (version 3.0.2; audEERING GmbH [43] with the emobase and emobase2010 configurations, yielding a 570D feature vector for each utterance. The feature space comprised 30 low-level descriptors summarized by 19 functionals, covering intensity and loudness, spectral properties (eg, MFCCs), pitch-related measures (F0 and derived measures), voice quality indexes (jitter, shimmer, and harmonic-to-noise ratio [HNR]), and short-term signal characteristics (eg, zero-crossing rate). Full feature definitions and naming conventions are provided in Multimedia Appendix 1.

All analyses were implemented in R (version 4.3.3; R Foundation for Statistical Computing). To prevent optimistic bias due to within-participant leakage (ie, multiple recordings per participant), we used participant-level stratified 5-fold cross-validation, assigning all 16 utterances from the same participant to a single fold [30,31]. Stratification was performed by case-control status and sex to preserve balance across folds. Age was evaluated as a potential confounder; given the narrow age range and the focus on acoustic-only feature sets, models were primarily trained on speech features.

Within each training fold, we applied the following preprocessing steps: (1) removal of constant features, (2) elimination of highly correlated feature pairs (|r|≥0.90), and (3) minimum-maximum normalization using parameters derived from the training data and applied unchanged to the corresponding test fold. All data-driven steps—including normalization, feature filtering, feature selection, and any hyperparameter tuning—were performed using training data only.

All performance metrics were computed at the participant level (N=96). Participant-level predictions were obtained by averaging the 16 utterance-level feature vectors recorded for each participant to form a single participant-level feature representation, which was then used as input to the model. Accordingly, each confusion matrix adds up to 96.

Feature selection was conducted within each training fold using a prespecified mixed-ranking strategy integrating 3 criteria: random forest (RF) variable importance (50%), feature-outcome correlation (30%), and ANOVA statistics (20%) [32,33]. Features were ranked by aggregated scores, and the top 16 features were retained for model training in each fold. The selected features predominantly captured prosodic, spectral, and voice quality characteristics (eg, MFCC, Line Spectrum Pairs (LSP), F0 and derived measures, and HNR).

We evaluated 7 machine learning algorithms: RF, elastic net generalized linear model (GLMNET), Extreme Gradient Boosting (XGB), support vector machine (SVM; radial basis function kernel), k-nearest neighbor (KNN), naive Bayes (NB), and a feed-forward neural network (NNET).

Given the modest sample size (N=96), hyperparameter settings were intentionally conservative to reduce overfitting. Analyses were conducted using standard R packages: randomForest (RF), glmnet (GLMNET), xgboost (XGB), e1071 (SVM), caret (KNN and NNET), and naivebayes (NB). Hyperparameter settings were as follows: RF used 300 trees with a minimum node size of 10; GLMNET used an α of 0.5 with λ selected via inner cross-validation; XGB used max_depth of 3, learning_rate of 0.1, and a subsample of 0.8 with L1 and L2 regularization (α=0.5; λ=1); SVM used a radial basis function kernel with C of 1; KNN and NNET were fit via caret with limited inner resampling for parameter selection; and NB used default parameters without tuning. All tuning and model fitting was confined to training folds.

Model discrimination was assessed using AUC as the primary metric alongside accuracy, sensitivity, specificity, precision, negative predictive value, and F₁-score. Receiver operating characteristic curves and confusion matrices were used to summarize discriminative performance and error patterns. Metrics were defined as follows:

Feature stability was quantified by the frequency with which each feature was selected across cross-validation folds. Features selected in 80% or more of folds were classified as stable [34]. All preprocessing parameters were derived from training data and applied identically to test data to ensure reproducibility.

The final sample comprised 96 participants (n=48, 50% per group) aged 16 to 25 years (mean age 18.76, SD 1.48 years), with 42.7% (n=41) female participants overall. Gender distributions were comparable between groups (χ²₁=0.17; P=.68). Age differed between groups, with the SI group being older than the control group (SI: mean 19.25, SD 1.78 years; controls: mean 18.27, SD 0.92 years; t₉₄=−3.39; P=.001; Table 1).

As shown in Table 2, under participant-level, leakage-aware stratified 5-fold cross-validation, 6 of the 7 models achieved acceptable discrimination (AUCs>0.70), whereas NNET performed less well (AUC=0.629). RF yielded the best overall performance (AUC=0.813; accuracy=0.748; sensitivity=0.730; specificity=0.782). NB ranked second (AUC=0.806) and achieved the highest accuracy (0.757). XGB (AUC=0.781) and SVM (AUC=0.772) also showed good discrimination. KNN achieved the highest specificity (0.855) but lower sensitivity (0.645), whereas GLMNET showed more modest discrimination (AUC=0.729). Receiver operating characteristic curves and confusion matrices are shown in Figures 1 and 2, respectively.

Beyond aggregate metrics, an analysis of the confusion matrices (Figure 2) revealed specific patterns of misclassification. For the best-performing model (RF), while it correctly identified 72.9% (35/48) of the SI cases, 27.1% (13/48) of the participants with SI were incorrectly classified as controls (false negatives). Conversely, 20.8% (10/48) of the controls were misclassified as SI (false positives). These results indicate that, while the model captures core acoustic markers of distress, 27.1% (13/48) of SI cases remained undetected, highlighting the current limitations of a purely acoustic approach in this cohort.

The 20 most frequently selected features across cross-validation folds are shown in Figure S1 in Multimedia Appendix 1. Eight features were selected in at least 4 folds, indicating relatively consistent selection across resamples. These features predominantly reflected pitch variability (F0-related indexes), spectral shape (MFCC kurtosis), and harmonicity (HNR; Tables S2-S4 in Multimedia Appendix 1).

To characterize group differences for the robust feature set, we conducted independent-sample 2-tailed t tests and correlation analyses with false discovery rate (FDR) correction (Tables S2-S4 in Multimedia Appendix 1). After FDR correction, 7 of the 8 robust features remained significantly associated with SI status. Compared with the control group, the SI group showed higher values in 6 features—F0_sma_iqr2-3, F0env_sma_stddev, F0env_sma_iqr1-2, F0env_sma_linregerrQ, mfcc_sma [5]_kurtosis, and mfcc_sma [12]_kurtosis—whereas HNR_sma_iqr1-2 was lower in the SI group. Effect sizes were in the medium to large range (Cohen d=0.44‐0.85). Correlation analyses showed a consistent pattern: the 6 F0 variability and MFCC kurtosis features correlated positively with SI status (r=0.28‐0.40), whereas HNR_sma_iqr1-2 was negatively correlated (r=−0.22). HNR_sma_stddev did not survive FDR correction.

Logistic regression analyses corroborated the univariate findings. Higher values of F0 variability indexes and MFCC kurtosis were associated with increased odds of SI group membership, whereas HNR_sma_iqr1-2 was associated with decreased odds. To facilitate interpretability, standardized odds ratios (ORs; per 1 SD increase) were reported: MFCC kurtosis features showed standardized ORs of 1.94 and 2.25, standardized ORs for F0-related features ranged from 2.03 to 2.50, and HNR_sma_iqr1-2 showed a standardized OR of 0.62 (Tables S2-S4 in Multimedia Appendix 1).

Mixed-effects models further indicated that sentence type accounted for substantial variance in most acoustic features (Figure S2 in Multimedia Appendix 1). Significant group effects were observed for mfcc_sma [5]_kurtosis (F_1,94=9.048; P=.003) and mfcc_sma [12]_kurtosis (F_1,94=7.718; P=.007), with nonsignificant sentence type effects and interactions (P>.05 in all cases). F0-related features showed both group effects (F_1,94=10.296‐17.373; P<.001 to P=.002) and sentence type effects (F_3,1434=13.256‐75.388; P<.001) without evidence of group × sentence type interactions. For HNR features, sentence type effects were pronounced (F_3,1434=5.684 and 216.477; P<.001), whereas group effects were weak or marginal (HNR_sma_stddev: F_1,94=2.593 and P=.11; HNR_sma_iqr1-2: F_1,94=4.664 and P=.03), and interactions were nonsignificant.

To interpret model behavior, we examined RF feature importance and visualized marginal effects using partial dependence plots. F0 variability features ranked among the most influential predictors (Figure S3 in Multimedia Appendix 1). Partial dependence plots suggested nonlinear relationships between feature values and predicted SI risk for several features (Figure S4 in Multimedia Appendix 1), whereas HNR_sma_iqr1-2 showed a nonmonotonic pattern.

Shapley additive explanations analyses provided convergent evidence regarding the prominence of the robust feature set (Figure S5 in Multimedia Appendix 1; Figure 3). F0_sma_iqr2-3 and F0env_sma_stddev generally contributed positively to SI predictions, whereas HNR-related features tended to contribute negatively. Notably, mfcc_sma [12]_kurtosis showed bidirectional Shapley additive explanations values (−0.087 to 0.174), and F0env_sma_stddev exhibited the largest spread (−0.100 to 0.196), indicating heterogeneity in feature contributions across individuals.

This study presents a methodologically rigorous proof-of-concept framework for detecting SI from standardized, smartphone-recorded read speech in university students. Speech offers a low-burden complementary behavioral signal that may mitigate reliance on self-disclosure in SI assessment [35]. The central contribution is a leakage-resistant modeling and validation pipeline built around (1) standardized prompts, (2) explicit reading compliance checks, (3) participant-level cross-validation to prevent within-speaker contamination, and (4) transparent feature stability and interpretability analyses. These design choices directly address key validity threats in speech-based modeling, particularly speaker leakage and evaluation bias.

While our results are promising, the interpretation must be matched to our specific methodology and sample characteristics. Notably, the SI group was significantly older than the control group (P=.001). Although this age difference was numerically small (approximately 1 year) and occurred within a narrow developmental window (late adolescence), it represents a potential confounding factor that limits the direct attribution of all acoustic variance to psychological status alone. Furthermore, our use of a constrained read speech task, while it improved standardization, may have limited the expression of more intense, spontaneous emotional signals often captured in crisis hotline studies.

The best-performing model in terms of AUC was RF (AUC=0.813; accuracy=0.748), followed by NB (AUC=0.806; accuracy=0.757). The observed misclassification rate—particularly the false negatives—merits clinical consideration. Misclassification in SI detection may stem from several factors: (1) the “silent” or “stoic” phenotype of ideation where vocal prosody remains intact despite internal distress; (2) individual variability in baseline speaking styles that overlaps with SI-linked features; and (3) the modest sample size (N=96), which may not capture the full heterogeneity of vocal expressions in SI. Consequently, our model should be viewed as a screening aid rather than a definitive diagnostic tool, consistent with the need for cautious deployment of AI in mental health.

These results fall within the performance range reported in several influential speech-suicide machine learning studies, although differences in population, speech task, labels, and validation design constrain direct numeric comparisons. Su et al [22] analyzed crisis hotline calls (naturalistic, emotionally loaded speech with clinician-rated risk labels) and reported the best model performance in terms of accuracy at approximately 0.75, with F₁-scores of approximately 0.70. Belouali et al [36] used mobile app narrative recordings in US veterans and reported RF performance of approximately 0.80 (AUC) with sensitivity of 86% and specificity of 70% when combining acoustic and linguistic features. Min et al [21] evaluated voice from clinical interviews in patients with mood disorders and observed more modest between-person discrimination (AUC≈0.62) but improved within-person prediction of worsening suicidality over time (accuracy≈0.79; AUC≈0.67), highlighting how label definition (cross-sectional discrimination vs longitudinal change) and recording context can shift achievable performance. Pillai et al [37] further demonstrated that cross-dataset generalization is frequently poor even when within-dataset metrics appear strong, indicating that distribution shift must be considered alongside algorithm choice.

Compared with the standardized reading tasks widely used in suicide prediction research based on acoustic features, such as the rainbow passage [38,39], our standardized reading materials incorporated a wider variety of sentence structures and sentence functions. The sentences were relatively more independent from one another and exhibited greater variability across items. Furthermore, compared with the reading passages commonly used in English-speaking contexts, our materials were better suited to Chinese phonological characteristics, thereby providing a standardized reading task for populations with Chinese linguistic backgrounds.

In this context, the observation that a brief read speech task—more constrained than narratives or crisis calls—still yielded discrimination comparable to that of several naturalistic speech approaches suggests that standardization can retain clinically relevant acoustic variability while improving controllability and reproducibility [15].

A major methodological strength of the present work is the explicit handling of within-participant leakage introduced by multiple recordings per speaker (16 utterances per participant). Leakage is a known failure mode in speech machine learning: if folds are split by recordings rather than participants, models can inadvertently learn speaker identity or stable vocal traits, inflating apparent performance [40]. Participant-level stratified cross-validation—where all utterances from a participant remain in one fold—directly addressed this issue and is consistent with best practice recommendations in behavioral machine learning and speech-based mental health research. In addition, the compliance pipeline (automatic transcription plus fuzzy alignment and manual review) reduced noise from off-prompt or incomplete readings, which can otherwise obscure acoustic signal in smartphone audio collected outside laboratory conditions. Such quality control procedures are often underdescribed in the literature but are likely to become increasingly important as speech data collection shifts toward real-world mobile settings [36].

The feature patterns most consistently associated with SI—indexes of pitch variability (F0-related measures), spectral shape (MFCC kurtosis), and harmonicity (HNR)—are broadly consistent with prior work linking suicidality and related affective states to altered prosody and voice quality while also illustrating how different speech tasks may privilege different acoustic cues. In hotline settings, spontaneous emotional speech likely amplifies prosodic instability and arousal-linked voice changes, whereas standardized reading reduces linguistic variability and may highlight subtler motor or phonatory correlates [22]. In narrative recordings, acoustic and linguistic features jointly contribute to psychiatric classification, suggesting that, when free speech is available, content-related information can be additive; standardized prompts isolate acoustics, making inferred associations more interpretable as vocal correlates rather than mixed vocal-semantic signals [25,41,42]. Reviews of speech processing in depression and suicidality similarly emphasize recurrent involvement of pitch, energy, spectral measures (often MFCC derived), and voice quality features while noting that effect direction and magnitude can vary by task, language, and recording conditions [15].

The comparability of performance estimates across studies depends not only on metrics but also on sampling strategies and outcome construction. Hotline and telehealth studies often involve higher base rates of distress and labels reflecting clinician-rated risk strata rather than strictly defined ideation thresholds, which can influence separability and reported metrics [22]. Conversely, balanced case-control designs improve statistical efficiency for early-stage method development but depart from community prevalence distributions. Consequently, accuracy in a balanced sample should not be interpreted as an expected positive predictive value for population screening without recalibration [43]. These considerations underscore the importance of transparent reporting and leakage-aware validation, and they help explain why methodological heterogeneity in recording contexts and analytic pipelines continues to challenge transportability across settings [20].

Methodologically, the pattern observed in this study—strong performance of simpler, well-regularized models (RF and NB) alongside weaker performance of the neural network—aligns with prior observations that more complex architectures can overfit in small, heterogeneous datasets without extensive harmonization [21]. Consistent with this, cross-dataset evidence indicates that robustness under distribution shift remains a central bottleneck even when within-dataset metrics appear strong, highlighting the need for evaluation designs and dataset diversity that explicitly test generalizability [37].

In summary, the present work contributes a reproducible methodological template for SI detection from voice by combining standardized prompts, compliance quality control, leakage-aware participant-level validation, and transparent feature stability analyses. When viewed alongside hotline, clinical interview, and narrative speech studies, these results support the feasibility of extracting suicidality-relevant information from speech across diverse contexts while underscoring the importance of rigorous evaluation and cross-context testing as the next step for the field.

Several limitations merit consideration. First, the cross-sectional nature and specific recruitment from a single university limit the generalizability of our findings. The identified acoustic markers should be interpreted as correlational signals rather than markers of future suicide risk. As highlighted by the issue of misclassification in our sample, these features alone are insufficient for high-stakes clinical decision-making without integration into a broader, multimodal assessment framework. Second, while the standardized read speech task controlled linguistic content and reduced between-speaker variability, it may have constrained the range of emotional and expressive speech characteristics captured. Accordingly, the ecological validity of the identified acoustic patterns may be lower than that achievable with spontaneous or emotion-elicited speech. Third, the neural network–based model showed weaker performance than the simpler models in this dataset, plausibly reflecting the moderate sample size and the challenges that complex models face in learning stable, generalizable representations from heterogeneous speech data. Fourth, our pipeline relies exclusively on acoustic features derived from standardized read speech, ignoring linguistic content and physiological biomarkers. These limitations are consistent with concerns raised in recent reviews and empirical work on acoustic-based suicide risk assessment [20].

Future research should strengthen the construct validity of acoustic indicators and evaluate their sensitivity to individual differences in SI. First, expanding recruitment to multiple universities and larger samples will improve statistical power and enable testing of whether the identified acoustic features consistently reflect underlying psychological constructs—such as emotional dysregulation, cognitive constriction, or motivational withdrawal—across diverse student populations. Second, future studies should explicitly model individual differences in speech production and psychological functioning. Beyond group-level classification, person-centered and longitudinal designs may help disentangle stable, trait-like vocal characteristics from state-dependent fluctuations associated with changes in suicidal risk, thereby improving interpretability and reducing misclassification driven by normative speaking style variability. Third, efforts to develop early warning systems should remain grounded in psychological theory and measurement principles. Acoustic features are best evaluated as indicators within a multi-method assessment framework, complementing self-report and clinical judgment rather than serving as stand-alone predictors. Aligning model outputs with interpretable psychological constructs may facilitate responsible integration into prevention contexts while preserving conceptual clarity and ethical safeguards. Fourth, moving beyond the unimodal acoustic approach, future research should explore multimodal fusion that integrates speech acoustics with natural language processing and physiological biomarkers. Multimodal fusion may be a superior method for detecting SI as it captures complementary risk dimensions—such as lexical markers of hopelessness, prosodic instability, and autonomic arousal—while compensating for the limitations of any single modality.

This study demonstrates the feasibility of using smartphone-acquired acoustic features from standardized read speech to discriminate SI among Chinese university students. The RF model showed promising preliminary performance (AUC=0.813) and identified stable acoustic markers. These findings support speech-derived acoustics as a low-burden, objective complement to traditional self-report screening in campus mental health contexts, where concealment and resource constraints can limit detection. Importantly, acoustic indicators should be interpreted as correlational signals rather than diagnostic tools, and their use should be embedded within existing counseling pathways and ethical safeguards. With further external validation and evaluation in more diverse, real-world settings, speech-based acoustic analysis can potentially alleviate the burden on campus counseling centers by providing an automated, preliminary triage stage in a multimodal screening framework.