0
0
New Delhi’s premier medical institute is testing whether subtle changes in everyday speech could help flag depression earlier and more objectively than traditional questionnaires alone.
What the AIIMS study found
Researchers at the All India Institute of Medical Sciences (AIIMS), New Delhi, have analysed voice recordings from 423 participants at a specialised Speech Health Lab on campus. Most participants were young adults, with a mean age of about 24 years and roughly three-quarters under 25. Standard psychiatric screening tools showed that around 32% of these individuals had clinically meaningful depressive symptoms.
When researchers ran participants’ voice samples through automated speech-analysis systems, the models could distinguish those with depressive symptoms from others with an accuracy ranging from about 60% to 75%, reaching nearly 78% when longer speech samples were used. The algorithms examined both what people said (linguistic features such as fluency and articulation) and how they said it (paralinguistic features such as tone, pitch, emotional “colouring,” and vocal energy).
The AIIMS team observed that depression was often associated with reduced fluency, flattened prosody (more monotonous speech) and lower vocal effort, patterns that align with broader research on “vocal biomarkers” of mood disorders.
Why voice is emerging as a “digital biomarker”
A growing body of research suggests that depression affects cognitive and emotional processes that in turn influence speech production, including rhythm, pitch, loudness and pauses. A 2025 systematic review of artificial intelligence–driven speech and voice analysis for depression reported that voice-based models could distinguish people with depression from healthy controls with an area under the curve (AUC) of about 0.71 to 0.93 in multiple studies, indicating moderate to high diagnostic performance.
In one cross-sectional and longitudinal study, a neural network trained on 30 acoustic voice features predicted depression severity scores with a correlation coefficient of 0.68 and mean absolute error of just over 3 points on a standard depression scale. Some acoustic features, including measures related to loudness and specific spectral characteristics, also changed in tandem with improvement during psychotherapy, suggesting potential use for monitoring treatment response
Recent work in digital mental health has further shown that speech features such as pitch variability, speech rate and the length of pauses can reflect momentary shifts in mood and energy in daily life, not just static clinical diagnoses. Collectively, these findings support the idea of voice as a non-invasive, scalable “digital biomarker” that could complement traditional psychiatric assessments rather than replace them.
The wider mental health context in India and globally
Depression remains one of the leading causes of illness and disability worldwide. A recent analysis using Global Burden of Disease data estimated that prevalent cases of depressive disorders increased by about 88% globally between 1990 and 2021, with more than 330 million people currently affected. The World Health Organization (WHO) estimates that roughly 4% of the global population experiences depression at any point in time, with higher rates among women and older adults.
In India, the National Mental Health Survey (2015–2016) found that around one in 20 adults lives with a depressive disorder, and highlighted suicide as a major associated risk. A separate study of 8,542 college students across 15 Indian cities by the National Institute of Mental Health and Neurosciences (NIMHANS) reported that nearly one-third had moderate to severe depressive symptoms and close to one in five had experienced suicidal thoughts.
Despite this burden, treatment gaps remain wide. WHO estimates that in many low- and middle-income countries, 76–85% of people with mental disorders receive no treatment, owing to factors such as limited specialist availability, affordability barriers, and stigma. This treatment gap has driven interest in low-cost, technology-enabled tools that can support early detection and triage in primary care and community settings.
What experts are saying
Dr Nand Kumar, professor in the department of psychiatry at AIIMS Delhi, emphasised that speech analysis can provide an objective window into the cognitive and behavioural changes associated with depression. “Patients with depression often show reduced fluency, diminished prosody or monotonous speech,” he noted, adding that these changes can be captured and quantified by automated systems.
Independent experts, not involved in the AIIMS project, echo the potential but urge caution. Researchers in voice-based mental health tools stress that machine-learning models must be rigorously validated, transparent and tested across diverse languages, dialects and cultural contexts to avoid bias and misclassification. Methodological work has highlighted that how voice data are collected (for example, scripted tasks versus spontaneous conversation, clinical settings versus mobile apps) can significantly affect performance and generalisability of depression-detection algorithms.
Clinical psychiatrists also underline that depression is a complex condition influenced by biological, psychological and social factors, and that any algorithmic assessment should be interpreted within a broader clinical context. Voice biomarkers, they say, might eventually function like an additional vital sign that prompts further evaluation, rather than a stand-alone diagnostic decision-maker.
Implications for screening, care and everyday life
If further validated, voice-based tools could help health systems scale up mental health screening, particularly in resource-constrained settings where access to psychiatrists and psychologists is limited. In primary care, a short spoken task captured on a tablet or smartphone could, in theory, provide an initial risk signal, prompting the clinician to administer a more detailed assessment or make a mental health referral.
For younger adults—who were heavily represented in the AIIMS study—integrating speech-based checks into digital platforms they already use, such as helplines, teleconsultations or mental health apps, could lower the threshold for seeking help. For example, periodic check-ins where users speak for a minute or two about their day could generate risk scores over time, alerting users or clinicians when patterns suggest worsening mood.
However, experts caution that these tools should never be used for self-diagnosis or as a substitute for a comprehensive mental health evaluation. A low-risk reading does not rule out depression, and a high-risk reading does not confirm it; both are prompts to have a conversation with a qualified professional. For individuals, the practical takeaway is to view such technologies as potential aids in monitoring emotional health—much like a fitness tracker for mood—rather than as definitive arbiters of diagnosis.
Limitations, open questions and safeguards
While the AIIMS findings are promising, several important limitations remain. The current analysis is based on a single-centre sample of 423 participants, predominantly young adults, which may not fully represent older adults, rural populations or people speaking different Indian languages and dialects. Reported prediction accuracies between 60% and 75% (up to 78% with longer recordings) are useful for screening but fall short of what would be required for a stand-alone diagnostic tool.
Systematic reviews of voice biomarkers for depression note considerable variation in study design, sample size, recording conditions and validation methods, with several studies judged to have a substantial risk of bias. Many models are trained on relatively small or homogeneous datasets, raising concerns about overfitting and reduced performance in real-world environments. In addition, emotional states such as stress, anxiety, fatigue or physical illnesses affecting the throat or lungs can also alter voice, potentially confounding depression predictions.
Privacy and ethical safeguards are another major consideration. Voice samples are inherently identifiable and can carry sensitive information, so experts stress the importance of robust data protection, clear consent processes, and strict limits on how recordings are stored, analysed and shared. Regulators and professional bodies are beginning to explore frameworks for evaluating and certifying AI-based mental health tools, but formal guidelines specific to voice biomarkers are still emerging.
What this means for readers now
For readers, the message is twofold. First, if you notice changes in your own speech—speaking much more slowly, struggling to find words, or sounding unusually flat or monotonous over weeks—it can be one among several cues to check in on your mental health and talk to a trusted professional or helpline, especially if accompanied by low mood, loss of interest or thoughts of self-harm. Second, as speech-based screening tools start appearing in clinics, apps or helplines, they should be viewed as early-warning supports, not as clinical verdicts.
Health systems in India and worldwide are exploring how to integrate such technologies into broader mental health strategies that also include public education, suicide prevention, counselling services and access to affordable treatment. For now, the AIIMS research adds to a growing evidence base suggesting that our voices may carry valuable information about our emotional state—and that with careful, ethical use, technology could help clinicians listen a little more closely.
Medical Disclaimer: This article is for informational purposes only and should not be considered medical advice. Always consult with qualified healthcare professionals before making any health-related decisions or changes to your treatment plan. The information presented here is based on current research and expert opinions, which may evolve as new evidence emerges.
References
-
AIIMS Delhi Speech Health Lab preliminary findings on speech analysis and depression. “Depression Detectable In Voice: AIIMS Researchers.” ETHealthworld / The Economic Times, 31 January 2026.[economictimes]
