Automated Behavioral Health Assessment Using Vocal Biomarkers

This peer-reviewed study, accepted to the 2026 Proceedings of Artificial Intelligence in Medicine (AIME), introduces a voice-based clinical decision support tool that screens for a generalized behavioral health condition (Major Depressive Disorder (MDD), Generalized Anxiety Disorder (GAD), or both) from just 60 seconds of spontaneous speech. Rather than relying on what a person says, the model analyzes acoustic features of the voice, making it language and topic-independent and applicable across in-clinic and remote settings. Using a demographically diverse dataset of speech responses to a simple "How are you?" prompt, labeled against the validated PHQ-8 and GAD-7 instruments, the researchers compared Random Forest, XGBoost, and deep neural network classifiers built on X-vector, Wav2Vec2, TRILLsson, and HuBERT embeddings. To strengthen clinical safety, the system also introduces an "uncertain" classification category that flags low-confidence predictions near the decision boundary rather than forcing a binary result.

The best-performing ensemble models achieved a UAR of 0.70 for depression and 0.68 for anxiety, with the combined behavioral health assessment reaching a sensitivity of 0.76 and specificity of 0.65 on the remote test set. Performance held up, and even improved, on an independent in-clinic dataset collected via tablet, where the model achieved a sensitivity of 0.68 and specificity of 0.80. These results demonstrate that a consolidated vocal biomarker can serve as a scalable, non-invasive screening mechanism to identify individuals at risk for depression and anxiety, two of the most prevalent and undertreated mental health conditions worldwide (affecting roughly 332 million and 359 million people globally, according to the WHO). By surfacing high-risk individuals quickly and reliably from a brief voice sample, this approach offers a practical path to earlier intervention and broader access to behavioral health care.

Reference: Namhee Kown, Nate Blaylock, Raymond Brueckner, Vinod Subramanian, and Henry O'Connell. Automated behavioral heatlh assessment using vocal biomarkers, Proceedings of Artificial Intelligence in Medicine. 2026.