Science1 publisher3 min readPublished
A speech classifier scored 86.2% on patients psychiatrists had already diagnosed
A team in the Netherlands trained software on 88 acoustic features from diagnosed patients and healthy controls. The year-and-a-half diagnostic delay it is pitched at happens earlier, among undiagnosed patients.
The Scientist · Science desk

What happened
- Americans with psychotic disorders, more than three million of whom have schizophrenia, receive a diagnosis on average a year and a half after their first symptoms appear.
- A team in the Netherlands measured 88 speech features, among them loudness, pause length, vowel pronunciation and intonation, in recordings of people psychiatrists had already diagnosed.
- Tested on audio from patients it had not heard before, the software told diagnosed patients from healthy controls with 86.2 percent accuracy and separated illness subtypes.
- Diagnosis today runs through symptom rating scales filled in by clinicians, and two clinicians scoring the same patient can land 30 to 50 percent apart.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint A model validated on diagnosed patients against healthy controls still cannot answer the clinician's question about a first-episode 20-year-old: schizophrenia, or something else. That judgment is where most of the 18 months goes.
- decision Because the analysis runs on a few minutes of recorded conversation, a service that wants it has to decide first about recording psychiatric appointments, retention periods and patient consent.
- exposure Anyone buying speech-diagnosis software on a headline accuracy number is exposed to noise in the reference standard, since the labels come from a scoring process that varies 30 to 50 percent between raters.
- contradiction Insel frames the technology as measurement psychiatry has never had, while the study coauthor says it adds least where diagnosis is already clear; the two positions point at very different readiness for use.
The training design sets the limit on what 86.2 percent can support. On one side were recordings of people psychiatrists had already diagnosed; on the other, healthy controls [4][5]. Both ends of that comparison carry confident labels, and the problem has only two groups. The hard case in a clinic is different. A clinician is weighing schizophrenia against another mental illness, and does it by tracking which symptoms show up over time [10].
What the software measures sits downstream of an old clinical observation. Patients can sound more monotonous or robotic, take longer pauses, and lack the volume contrast healthy speakers use, and psychiatrists treat speech as a marker of disordered thinking, a hallmark of the illness [15][16]. Those acoustic differences are hard to use for diagnosis in the early stages [15].
Accuracy of 86.2 percent leaves 13.8 percentage points of error, roughly one classification in seven going the wrong way [6]. The direction matters more than the total, because a false positive in a 19-year-old and a missed case are different clinical events. The Scientific American account reports the single accuracy figure and does not break out sensitivity and specificity or say how many recordings sit behind it [5].
There is also the standard the model was graded against. Diagnoses come out of rating scales filled in from what a patient says and how they say it, and clinicians' scores for one patient can differ by 30 to 50 percent [3]. A model trained on those diagnoses learns psychiatrists' decisions, disagreements included [4].
Thomas Insel, the psychiatrist and neuroscientist who led the US National Institute of Mental Health for 13 years, put the appeal in measurement terms. "For the first time, we could have a way of saying objectively, how delusional is this? How loose are these associations? How incoherent is it?" he said [9].
Alban Voppel, a study coauthor now at McGill University in Montreal, said the approach helps least with the clearest cases, and that AI could help detect patients in early stages or at high risk of becoming schizophrenic [7]. The second half is a hypothesis; what was tested here was diagnosed patients against healthy controls [5].
The clinical case for moving earlier is well established: the longer schizophrenia goes untreated, the poorer the response to treatment and the greater the risk of brain tissue loss, worsening symptoms and suicide [11]. About 23 million people worldwide have it, and diagnosis usually arrives between the late teens and early 30s [12].
The input is a few minutes of conversation [13]. In a clinic that means a recording of a psychiatric appointment, kept long enough to be analyzed, and the clinic adopting it decides who holds that audio, for how long, and what the patient agrees to. Scientific American reported that the software will not be in clinics tomorrow [14].
What to watch
- A prospective study that records people at clinical high risk, then scores the classifier against who converts.
- Sensitivity and specificity published separately, with the number of recordings in each group.
- A test of schizophrenia speech against speech from patients with other psychiatric diagnoses, the comparison clinicians find hardest.