Depression is a widespread mental health disorder that often remains underdiagnosed due to limitations in traditional assessment methods. Recent technological advances have enabled researchers to explore speech and voice characteristics as potential biomarkers of mental disorders. Indeed, voice-based assessment is promising because speech patterns may reflect cognitive, emotional, and psychomotor changes associated with depression. However, the literature remains heterogeneous in terms of datasets, modeling techniques, and evaluation methods.
To further know how voice-based biomarkers are associated with depression, Donaghy et al. (2024) conducted a narrative review with a systematic search examining the use of machine learning (ML) assessing techniques. Furthermore, the authors analyzed existing empirical studies to evaluate the effectiveness of speech-based indicators in detecting depressive symptoms and to explore methodological differences across studies. Several methodological key aspects were considered for the narrative and systematic review:
- Identification of studies focusing on speech features and ML models for depression detection.
- Inclusion of empirical research examining voice-based datasets and computational modeling techniques.
- Extraction and comparison of methodological characteristics such as sample size, voice features, ML algorithms, and evaluation metrics.
What is the efficacy of voice biomarkers for depression identification? What are the variations across samples and design methodologies?
A total of 19 studies were included in the final review. Across studies, researchers extracted various acoustic features from speech recordings, including: a) prosody (e.g., pitch, intensity, speaking rate); b) spectral features; c) temporal characteristics of speech. These features are thought to reflect psychomotor slowing, reduced emotional expression, and cognitive changes commonly associated with depressive states.
Moreover, this review examined several ML algorithms which were used across the selected studies, such as: a) Support Vector Machines (SVM); b) Random Forest; c) neural networks; d) deep learning models. Results across studies indicated that ML models can classify depressive vs. non-depressive speech with moderate to high accuracy, though performance varied depending on dataset quality and modeling approach.
Finally, in terms of methodological variability, the authors found considerable heterogeneity in sample sizes, data collection protocols, feature extraction method, and evaluation strategies. This variability impedes creating direct comparisons across studies and, therefore, limits conclusions about the best-performing approaches.
Promising efficacy of ML techniques for depression detection
The findings of this narrative and systematic review suggest that voice biomarkers analyzed with ML techniques represent a promising approach for depression detection. In fact, speech analysis offers a non-invasive, scalable, and potentially deployable technique through smartphones or telehealth platforms. Hence, ML may become a powerful tool for clinical assessment.
Nevertheless, the authors identify several limitations across the reported studies, including small and non-diverse samples, the lack of standardized datasets, inconsistent reporting of model performance, and limited replication. These gaps might be overcome by using larger datasets, standardized methodologies, and assure cross-validation assessment across populations to improve reliability and clinical applicability.
In conclusion, this study concludes that ML-based voice biomarkers show strong potential as digital indicators of depression, offering opportunities for early detection and remote mental health monitoring. However, further research is necessary to improve methodological consistency and to validate models in real-world clinical settings before widespread adoption.
