🎙️ Many people with bipolar disorder are first told they have depression. A team from Keio University, Institute of Science Tokyo and Shizuoka University recorded patients in interviews, trained an AI on the audio, and found the strongest clues lay not in what people said but in how they sounded. The catch: the study only told patients apart from healthy volunteers.

Years of detours before the right diagnosis

Bipolar disorder brings repeated swings between mania, a state of elevated mood and energy, and depression. The World Health Organization estimates that 36 million people had the condition in 2023, about 0.4 percent of the world's population. On average, their lives are 13 years shorter than those of the general population.

The trouble often starts at the first appointment. Most people seek help when they are low, and at that point the manic side may not have shown itself yet. Prior research cited by Keio found that 76.8 percent were first diagnosed with depression, and that getting the right diagnosis took about six years on average. Another study, cited by Hamamatsu University School of Medicine in September 2026, put it at 8.7 years on average. Antidepressants, the standard treatment for depression, work less well in bipolar depression and may tip some patients into mania.

How you sound mattered more than what you said

Between 2017 and 2024, the team recorded 30- to 60-minute interviews at Keio University Hospital and collaborating sites across Japan (seven hospitals and three clinics). Besides free conversation, participants described three pictures and retold either Cinderella or Kaguya-hime, a Japanese folktale about a princess from the moon. The analysis covered 61 people with bipolar disorder and 99 healthy volunteers. The paper appeared in the Journal of Medical Internet Research on November 1, 2026 (US Eastern Time).

The AI was given two kinds of clues. One described sound: 3,473 features covering pitch, timbre and voice quality, with anything that depends on loudness stripped out. The other described content: 211 features on emotional words, word choice and sentence structure.

Separating patients from healthy volunteers, the model reached an AUC of 0.916. AUC is a score where 1 means perfect separation and 0.5 is no better than guessing. Sound alone scored 0.910. Content alone managed just 0.651. Between 96 and 98 percent of the features that mattered were acoustic.

Each mood state, whether depression, mania or mixed states, or euthymia (a stable stretch between episodes), was also compared with the healthy group, and AUC ranged from 0.805 to 0.874. Telling mood states apart within the patient group was much harder. Only depression versus euthymia produced a statistically significant result, with an AUC of 0.760.

The comparison that matters most is still ahead

The whole project started from the problem of bipolar disorder being mistaken for depression. Yet the study compared patients only with healthy people, and it excluded anyone who also had major depressive disorder, schizophrenia or certain other conditions. The paper itself lists the lack of comparison with other psychiatric disorders as a limitation.

The numbers come from validation within the same dataset; nobody has yet checked them on patients from other hospitals. The pool of candidate features is large relative to the number of people, which raises the risk of overfitting, where a model learns the quirks of its own data. Every patient was taking medication, and the authors write that they could not separate the euthymia result from the effect of antipsychotic drugs. All participants spoke Japanese.

In September 2026, a team led by Hamamatsu University School of Medicine analyzed facial expression, voice and gaze in remote interview videos, along with clinical information such as sleep. According to the university's announcement, among participants currently in a depressive episode it separated bipolar depression from unipolar depression at 79 percent accuracy and an AUC of 0.75 (the paper's abstract gives 78 percent and 0.76 overall). Those figures also came from validation within its own data. The designs are too different for a direct comparison, but it points to a harder question than telling patients from healthy people.

The Keio-led team says its next steps are comparisons with depression and other conditions, replication in languages other than Japanese, and validation on external data. Keio's announcement says such measures are not meant to replace clinical interviews and should serve only as an aid. According to the paper, Taishiro Kishimoto, a professor at Keio University, founded the company i2medical and has received fees from several companies.

"What if they use it in job interviews?"

The day after the announcement, Japanese social media showed worry alongside hope. On Togetter, a site that curates social media posts, a collection gathered posts fearing that employers might use the technique in hiring and screen out people with bipolar disorder. On Hatena Bookmark, a Japanese social bookmarking service, one commenter called it promising as a diagnostic aid but warned it could be used to analyze job applicants or livestreamers without their consent.

The universities had attached a caution of their own. Along with the paper, they released sample audio that recreates the voice features the study identified. The source voices were not patients' but came from healthy volunteers outside the study and other non-patient recordings. The announcement stresses that the features describe group-level trends, and that having such a voice in no way means a person has bipolar disorder.

Measuring the mind by voice, worldwide

In Denmark, a Copenhagen research team collected voice features from everyday smartphone calls and, in a 2021 study, compared 121 people with bipolar disorder against 38 healthy controls. The AUC was 0.76.

In the United States, Kintsugi, which built AI to pick up signs of depression and anxiety in speech, announced in February 2026 that it was shutting down its commercial operations. It had sought clearance through the FDA's De Novo pathway for new kinds of medical devices, but according to The Verge, it ran out of money while waiting on the final stage of its application. Its models were released to the public. In Japan, a 2022 report described MIMOSYS, an app from the company PST that turns a short spoken sample into a mood reading it calls "genki-atsu," roughly "vitality pressure." On the regulatory side, Japan's government has asked the public to flag rules that hold back AI, including in healthcare.

How a voice sounds is harder to choose than what it says. Inside the clinic, that could give doctors a useful clue. Outside it, the same tool could be used to screen job applicants. Where would people in your country draw the line?

References