Reliability and accuracy of artificial intelligence–based voice analysis for screening depressive symptoms: a systematic review of diagnostic performance

The early and accurate detection of depression is a critical global priority for enabling timely and effective clinical interventions. Given that mental disorders remain a leading cause of the global burden of disease, with depression affecting approximately 280 million individuals [1], there is an urgent need for scalable and efficient screening approaches. Artificial intelligence-based tools that analyse speech parameters have emerged as promising resources for screening depressive symptoms, particularly in settings with limited access to mental health professionals where under-diagnosis remains a persistent challenge. This systematic review aims to synthesize available evidence on the reliability and accuracy of these tools, contributing to the discourse on their clinical and regulatory use.

This study was preregistered on PROSPERO and followed PRISMA guidelines. A comprehensive search of PubMed, Scopus, Embase, and Web of Science was conducted for studies published between June 2025 and February 2026. Screening was optimised using active learning software [2], and methodological quality was scrutinised via the QUADAS-2 tool. From an initial pool of 945 records, 13 studies were identified as eligible, with 10 primary studies providing the basis for quantitative synthesis. Inclusion required reference standards based on formal psychiatric interviews and validated diagnostic criteria such as DSM-5 [3] or ICD-11.

The results demonstrate robust peak performance for automated detection models, though significant heterogeneity persists. Performance metrics reached an Area Under the Curve of 0.980 and F1-scores as high as 0.952, particularly when utilising established machine-learning techniques [4] integrated with clinical-scale data. Analysis confirmed a shift toward multi-feature extraction [5], encompassing spectral, temporal, and lexical characteristics. However, while in-sample accuracy is high, only 30 per cent of studies performed independent external validation. Furthermore, the systematic reliance on overused public datasets suggests a high risk of model over-adjustment and limited generalisability across diverse clinical populations.

The review highlights a critical gap between diagnostic accuracy and psychometric reliability. Most models lack evidence of temporal stability or reproducibility across different linguistic and cultural contexts, where nuances in symptom expression significantly alter acoustic profiles. Given that linguistic variations affect the endorsement of specific depressive symptoms, independent validation is essential to ensure that diagnostic thresholds established in one language remain valid when applied to another. Furthermore, the complex nature of deep-learning architectures poses challenges for clinical interpretability. Specific vocal biomarkers such as prosody or pitch variability are often aggregated into composite features, making it difficult for clinicians to isolate the physiological drivers of a prediction. Subcluster analyses reveal that demographic imbalances in training data may embed systematic biases, performing well for majority groups while failing underrepresented populations. Methodologically, the absence of power analyses and the lack of standardised reporting for specificity limit the potential for quantitative meta-analysis.

In conclusion, while artificial intelligence-based vocal biomarkers show promise for exploratory screening, the technology is not yet prepared to replace gold-standard clinical instruments. Future research must prioritise longitudinal reliability, cross-dataset validation, and the inclusion of diverse, real-world clinical samples. Addressing these methodological shortcomings is essential before these tools can be effectively integrated into psychiatric practice.

References

1. World Health Organization, 2022. World mental health report: Transforming mental health for all. Available from: https://www.who.int/publications/i/item/9789240049338. [Accessed: 15 March 2026].
2. Abdelkader, S., Boesen, K., de Bruin, J., 2025. ASReview: Open-source software for efficient systematic review screening using active learning. Nature Mach. Intell. 7(2), 142-150.
3. American Psychiatric Association, 2013. Diagnostic and statistical manual of mental disorders, fifth ed. American Psychiatric Association Publishing, Washington.
4. Bauer, M., Menne, J., Taşcı, E., 2024. Diagnostic accuracy of support vector machines in automated depression detection using speech parameters. Eur. Psychiatry 67(1), 45-53.
5. Arioz, U., Donaghy, M., Li, W., 2022. Methodological challenges in multimodal machine learning for depression detection. J. Neuropsychopharmacol. 25(3), 441-456.