Validating objective and scalable speech markers of depression across two independent psychiatric cohorts

Abstract

Using speech as objective markers for major depressive disorder (MDD) has shown promise, yet their generalizability across clinical settings remains largely unvalidated. This study aimed to validate previously identified speech markers of depressive symptoms in an independent clinical cohort, thereby assessing their reproducibility and robustness for cross-site application. Speech data from two independent psychiatric cohorts (RWTH Aachen and University of Oldenburg, Germany) were analyzed, comprising 135 participants (71 healthy controls, 64 MDD patients). Participants completed a positive and a negative storytelling task, over 80 temporal, lexical, and spectral speech features were extracted from the acoustic signal. Statistical analyses assessed group differences and correlations with Beck Depression Inventory (BDI-II) scores. Machine learning models trained on the Aachen data were tested on the Oldenburg cohort. Several temporal and spectral speech features, including utterance duration, pause duration, and MFCCs, were consistently associated with MDD diagnosis and symptom severity across both cohorts. Machine learning models trained on Aachen data achieved a classification accuracy (ROC-AUC) of 0.63 on the Oldenburg sample, demonstrating above-chance but modest transfer performance. Voice quality features (shimmer, jitter) showed more variable associations: partial correlations indicated some significant effects (e.g., shimmer and jitter during positive storytelling), whereas moderation analyses revealed interaction effects, particularly for shimmer and jitter in negative storytelling, where MDD patients exhibited higher values in the Aachen cohort but lower values in the Oldenburg cohort compared to healthy controls. The study indicates that temporal and spectral markers of speech are relatively robust across independent clinical samples, whereas voice quality markers (shimmer, jitter) show site-dependent inconsistencies, acting as technical artifacts of varying recording conditions rather than robust biomarkers. While current speech-based classifiers remain less accurate than established self-report measures, their integration with clinical scores offers a more balanced trade-off between sensitivity and specificity. Future work should prioritize systematic evaluation across elicitation tasks, languages, and longitudinal settings to delineate which speech features are transferable and which are task-specific.