Currently accepted at: JMIR Formative Research
Date Submitted: May 8, 2026
Date Accepted: Nov 13, 2026
This paper has been accepted and is currently in production.
It will appear shortly on 10.2196/100735
The final accepted version (not copyedited yet) is in this tab.
Voice Markers for the Detection of Depressive, Anxiety, and Stress Symptoms among Dari/Farsi Speakers in Afghanistan: Cross-Sectional Model Development and Evaluation Study
ABSTRACT
Background:
Speech-based biomarkers offer a promising alternative for scalable and objective mental health screening, particularly in remote, care-deprived areas with limited access to mental health professionals.
Objective:
The present study developed and evaluated a multimodal speech-analysis pipeline integrating acoustic and lexical features for detecting depression, anxiety, and stress among Dari/Farsi-speaking adults in Afghanistan.
Methods:
A cross-sectional study was conducted with 263 Dari/Farsi-speaking adults in Afghanistan. Each participant completed the 21-item Depression, Anxiety and Stress Scale (DASS-21) and responded verbally to five open-ended questions. A three-stage pipeline was developed. Stage 1 (Acoustic) extracted handcrafted features using openSMILE eGeMAPSv02 and Praat, alongside self-supervised speech representations from XLS-R. Stage 2 (Lexical) derived text-based features from Whisper transcriptions using ParsBERT embeddings and TF-IDF with SVD. Stage 3 (Fusion) combined acoustic and lexical modalities through feature concatenation and probability-level ensembles. Logistic regression and multilayer perceptron classifiers were evaluated using five-fold nested cross-validation with strict group separation. Area under the curve (AUC) was the primary metric.
Results:
A total of 263 participants with 1,309 valid utterances were included. Multimodal fusion models outperformed unimodal approaches across all three outcomes. The best-performing models achieved AUC values of 0.793 (95% CI: 0.739–0.845) for depression, 0.775 (95% CI: 0.717–0.828) for anxiety, and 0.754 (95% CI: 0.693–0.809) for stress. All results were statistically significant (p < 0.001). Depression was the most detectable condition, followed by anxiety and stress. Handcrafted acoustic features remained competitive with self-supervised embeddings. TF-IDF-based lexical models sometimes outperformed contextual embeddings from ParsBERT.
Conclusions:
The present study provides the first evidence that a multimodal speech-analysis pipeline can detect depression, anxiety, and stress among a Dari/Farsi-speaking population. Integrating acoustic and lexical features consistently improved performance over single-modality approaches. In the short term, the pipeline offers a low-cost triage tool that could be embedded into primary care, community health worker, and mobile health workflows to extend screening to adults in remote, care-deprived districts of Afghanistan.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.