Voice biomarkers for depression use AI to analyze how people speak, such as pace, pitch and pauses, to flag possible depression. The largest published study found about 71% sensitivity and 74% specificity against a standard questionnaire. The signal is real, but performance often drops in new populations, and consent is the unresolved problem.
Depression changes how people sound. Clinicians have described it for as long as psychiatry has existed: slower speech, longer pauses, flatter intonation, a narrower range of pitch. What is new is the claim that software can hear these changes in 20 or 30 seconds of ordinary talk, and do it at scale, during a phone call to a clinic or a telehealth visit, without anyone filling in a form.
It is an appealing idea. Depression is common, underdiagnosed and often missed in busy primary care. A passive screen that runs in the background of calls that already happen sounds like exactly the kind of low-cost win health systems need.
The thesis here is more cautious. Voice biomarkers carry a genuine signal, but the evidence so far shows them roughly reproducing a questionnaire that costs nothing, with accuracy that tends to erode outside the populations they were trained on, and with a consent problem that grows as the technology becomes more invisible. The most important question is not whether voice AI can detect depression. It is whether it should be listening in the first place.
If you are feeling low or thinking about harming yourself, please contact your local emergency services or a crisis line. In the US you can call or text 988.
How accurate are voice biomarkers for depression?
The largest peer-reviewed study to date was published in the Annals of Family Medicine in January 2025 by Alexa Mazur and colleagues, evaluating a tool from the company Kintsugi. It analyzed at least 25 seconds of free-form speech from 14,898 adults in the United States and Canada, collected between 2021 and 2022, and compared the tool’s output with each person’s score on the PHQ-9 depression questionnaire, using a cut-off of 10 for moderate to severe symptoms.
In the validation group, the tool had a sensitivity of 71.3% and a specificity of 73.5%. Put simply, it caught about seven in ten people whose questionnaire suggested moderate or severe depression, and correctly cleared about three-quarters of those whose questionnaire did not.
That is a meaningful result for a signal extracted from tone rather than words. It is also worth reading carefully. The reference standard was not a psychiatric diagnosis. It was the PHQ-9, a free, nine-question self-report form that is itself a screening tool. The voice model was, in effect, trying to predict a questionnaire, and it did so imperfectly.
Kintsugi has described pursuing FDA review for its technology. At the time of writing, to our knowledge, no voice biomarker tool for depression has FDA marketing authorization as a diagnostic device.
What does 71% sensitivity mean in a real clinic?
Numbers like sensitivity and specificity are easy to nod at and hard to feel, so here is a hypothetical.
A primary care practice sees 1,000 adults, and suppose 15% of them, 150 people, would score 10 or more on the PHQ-9. A voice tool with the published accuracy would flag about 107 of them correctly and miss about 43. Among the 850 people without significant symptoms, it would wrongly flag about 225.
So the practice gets roughly 332 flags, of which about a third belong to people with moderate or severe symptoms. Every flag needs follow-up, which in practice means asking the person to complete the PHQ-9 or talk to a clinician. The practice could have simply given everyone the PHQ-9 in the first place, and many already do.
The case for voice tools therefore rests less on accuracy than on reach: they might screen people who would never fill in a form, such as patients calling a nurse line, or track change over time between visits without asking anyone to do anything. Those are plausible advantages. They have not yet been shown to change outcomes in trials.
Why do voice models struggle to generalize?
Voice is shaped by far more than mood. Accent, dialect, language, age, sex, dental work, a cold, the microphone, background noise and whether someone is reading or chatting all change the acoustic features a model relies on.
Research on speech-based depression detection has repeatedly found that models that perform well on one dataset degrade sharply when tested on another, particularly across languages and recording conditions. Some reviewers have also pointed to a methodological trap: when recordings from the same speaker end up in both training and test sets, a model can learn to recognize individuals rather than depression, inflating reported accuracy.
This matters for equity. If a model was trained mostly on English-speaking adults in North America, it may perform differently for older patients, non-native speakers, people with speech differences or those in other countries. As we discussed in Bias in Clinical Algorithms: The Fix That Matters Most Is the One We Rarely Do, the fix is local validation in the population where a tool will be used, which is the step most often skipped.
A practical test for any vendor is to ask for performance broken down by age group, sex, first language and recording channel, such as landline, mobile or video. If the answer is a single headline accuracy figure, the tool has not been validated for your patients. It has been validated for someone else’s.
Who consents when a voice is analyzed?
This is the hardest part, and the one most often treated as an afterthought.
A questionnaire is visible. You know you are being asked about your mood because you are reading the questions. A voice biomarker can run silently on a call about a prescription refill. The person may have agreed, somewhere in a set of terms, to calls being “recorded for quality and training purposes”. They almost certainly did not understand that to mean a mental health assessment.
Law is beginning to reach this. In Illinois, the Biometric Information Privacy Act treats voiceprints as biometric identifiers requiring informed written consent, and it has generated extensive litigation. In the European Union, the AI Act has since February 2025 prohibited AI systems that infer emotions in workplaces and schools, except for medical or safety reasons, a sign of how seriously regulators take inference from voice and face.
Health settings are not workplaces, and the medical exception exists for good reasons. But the principle travels. Inferring a person’s mental state from how they sound is a clinical act. It deserves the same consent as any other test, and the same right to decline. Voice data are also uniquely identifiable and hard to anonymize, so a recording kept for “model improvement” is a long-lived record of someone’s health.
There is an institutional question too: who else might want these signals? Insurers, employers and call centers all have commercial reasons to know who is depressed. Clear rules that voice-derived mental health inferences are used only for care, with the patient’s knowledge, need to exist before the technology is widespread, not after. We made a related argument about data rights in “Who Owns Patient Data?” Is the Wrong Question: what matters is who may use it and for what.
The strongest case for voice biomarkers
The optimistic argument deserves a fair statement. Depression is missed constantly, especially in men, older adults and people who do not describe emotional symptoms. Questionnaires rely on people answering honestly and on staff remembering to give them. Voice analysis is objective in the narrow sense that it does not depend on how a patient chooses to present, and it can be repeated effortlessly, potentially catching deterioration earlier.
As a monitoring tool with explicit consent, for example in a patient already in treatment who agrees to regular voice check-ins, voice biomarkers could add a useful data stream alongside the questionnaire, a little like how wearables add heart rate data to a cardiology follow-up. We explored that dynamic in Wearables Data Is Not Noise. It Is a Screening Programme Nobody Designed.
That is a reasonable vision. It is also a much narrower one than passive population screening, and it puts consent and clinical follow-up at the center rather than the edge.
What this means for health systems, clinicians and founders
For health systems, the sensible stance is to treat voice biomarkers as investigational. Pilot them only with explicit opt-in consent, validate locally, compare against the PHQ-9 you could already be using, and define in advance what happens to every flag.
For clinicians, a voice flag is at most a prompt to ask. It is never a diagnosis, and a negative result is not reassurance.
For founders, the business risk is not only accuracy. It is trust. A single story about an insurer or employer using voice-derived mental health data could set the whole field back. Building consent, data minimization and use restrictions into the product is a competitive advantage, not a compliance cost.
For regulators, the question is whether voice-based mental health inference outside clinical settings, in customer service, human resources or consumer apps, should face limits similar to the EU’s workplace ban.
Frequently asked questions
Can AI detect depression from your voice?
AI can detect patterns in speech associated with depression, such as slower pace and flatter tone. A 2025 Annals of Family Medicine study of about 15,000 adults found a voice tool identified around 71% of people with moderate to severe symptoms on a standard questionnaire. It is not a diagnosis, and a clinician should assess anyone who is concerned.
What are voice biomarkers?
Voice biomarkers are measurable features of speech, such as pitch, rhythm, pauses and energy, that may correlate with health conditions. Researchers are studying them for depression, anxiety, Parkinson’s disease and other conditions. Most remain experimental, and their accuracy varies with language, accent, age and recording quality.
Are voice biomarker tests FDA approved for depression?
As of early 2026, to our knowledge, no voice biomarker tool has FDA marketing authorization to diagnose depression. Some companies have described pursuing FDA review. Tools marketed for general wellness may be available without FDA authorization, but they are not validated diagnostic tests and should not be used in place of clinical assessment.
Is it legal to analyze someone’s voice for mental health?
It depends on the setting and jurisdiction. Illinois requires written consent to collect voiceprints under its biometric privacy law. The EU AI Act bans AI emotion inference in workplaces and schools except for medical or safety purposes. In health care, good practice is to seek explicit informed consent before any voice-based mental health analysis.
How is depression usually screened in primary care?
Most commonly with short validated questionnaires such as the PHQ-2 and PHQ-9, which ask about mood, interest, sleep, energy and other symptoms over the past two weeks. They are free and quick. A positive screen should lead to a clinical assessment, and anyone with thoughts of self-harm should seek help promptly.
The takeaway
Voice biomarkers have found a real signal in a very noisy place. Before they listen to millions of calls, the field needs to prove they add something a nine-question form does not, that they work for people who do not sound like the training data, and that every person being analyzed knows it and said yes.
Would you agree to have your voice analyzed for signs of depression during routine calls with your clinic? What would you need to know first?
