Speech Alone Predicts Schizophrenia Symptom Severity Across Ten Countries

Machine learning models trained on nothing but the sound of a person’s voice can estimate the severity of individual schizophrenia symptoms with clinically meaningful accuracy, according to the largest and most linguistically diverse study of its kind to date. The research, published in eClinicalMedicine, pooled speech recordings and clinical ratings from 453 patients with schizophrenia spectrum disorders across ten international sites, spanning languages from Czech and Turkish to Chinese Mandarin and Chilean Spanish. For each of eight symptom items from the Positive and Negative Syndrome Scale (PANSS), the best audio-only models achieved a root-mean-squared-error (RMSE) below 1.5 on a scale where each item is rated from 1 to 7 — an error margin the authors describe as practically useful at the item level.

The sites included the Czech Republic, the United States, Spain, Chile, France, Switzerland, the Netherlands, Türkiye, Germany and China.

Support vector regression performed best for conceptual disorganisation (segment-level RMSE 1.318), a two-layer neural network for delusions, social withdrawal and mannerisms, and a three-layer network for hallucinatory behaviour, blunted affect, lack of spontaneity and unusual thought content.

The choice of representation mattered. Statistical comparisons using the aligned rank transform showed that models built on m-HuBERT embeddings consistently outperformed those using concatenated features, which in turn beat models relying solely on hand-crafted acoustic-prosodic descriptors — with all pairwise differences statistically significant. Embeddings were best for six of the eight items; acoustic-prosodic features won only for social withdrawal and mannerisms. Prediction error was lowest for mannerisms and lack of spontaneity, moderate for blunted affect and unusual thought content, and highest for delusions and social withdrawal. Performance varied by language: French and Chinese recordings were consistently harder to predict, while Peninsular and Chilean Spanish, German and Dutch yielded lower errors.

The team also probed the models for demographic and task-related bias. Prediction errors showed no significant association with age or education, and no sex differences for seven of the eight items — the exception being social withdrawal, where errors were significantly larger for male participants. Errors did not differ significantly across elicitation tasks, which ranged from free speech and picture description to dream narration, reading and immediate recall. Symptom severity, however, told a different story: for all items except social withdrawal, higher observed symptom scores were significantly linked to larger prediction errors.

That severity gradient is a central caveat. The PANSS score distributions in the dataset were heavily imbalanced, with most observations clustered at lower symptom levels and the highest scores rare — a feasibility constraint the authors attribute to the difficulty of recruiting severely ill patients for speech-based studies. The models showed restricted prediction ranges and a systematic tendency to under-predict the upper end of the scale, meaning their validity is currently best supported for the lower-to-moderate symptom range. Predicting very high severity would require extrapolation beyond the training data, and the authors argue the models must be validated in high-severity cases before they can be considered reliable across the full clinical spectrum.

An exploratory out-of-sample test offered a preliminary glimpse of real-world generalisability. Five participants from an ongoing local study — two Spanish speakers and three Catalan speakers, the latter language absent from training — completed speech tasks and PANSS ratings at baseline and again roughly three months later. The models performed comparatively well on most items, with RMSEs including 0.666 for conceptual disorganisation and 0.352 for mannerisms, though hallucinatory behaviour fared poorly at 3.192. With only five people, the authors stress, this deployment should be viewed as a small use case rather than evidence of external validity.

Several further limitations temper the findings. The dataset was heterogeneous in languages, symptom assessments and elicitation tasks, and language differences were confounded with site differences, so the study cannot isolate which language-specific factors affect predictions. PANSS ratings themselves carry substantial inter-rater variability across sites, which the retrospective design could not standardise — noise that likely depresses measurable performance. The researchers also note that racial and ethnic bias in AI-based language models remains an open concern, that participant race and ethnicity data were not collected, and that the predictors remain largely black boxes, with detailed interpretability analysis deferred until robustness is better established.

Even so, the implications are considerable. Speech can be collected passively, remotely and cheaply, and analysed automatically at low marginal cost — properties that fit a clinical need for scalable monitoring of the symptom fluctuations that precede relapse. The authors suggest that reduced precision at very high symptom levels may matter less in practice, since marked exacerbations are usually clinically obvious, whereas small shifts in the lower-to-moderate range, where the models perform best, could provide earlier warning signals. The trained models have been released publicly, and the authors say the multilingual corpus supports transfer learning from high-resource languages such as English to low-resource ones such as Czech — a step toward including previously underrepresented languages and populations in computational psychiatry.

Subject of Research: A multilingual machine learning study predicting schizophrenia symptom severity from speech recordings alone.

Article Title: Predicting PANSS symptoms in schizophrenia spectrum disorders using speech only: an international, multi-centre, computational secondary analysis across multiple languages

Article References: He, R., Kirdun, M., Palominos, C., Orejudo, L. N., Barthelemy, S., Bhola, S., Ciampelli, S., Decker, A., Demirlek, C., Fusaroli, R., García-Molina, J. T., Gimenez, G., Hüppi, R., Koelkebeck, K., Lecomte, A., Qiu, R., Simonsen, A., Tourneur, V., Verim, B., … Hinzen, W. (2026). Predicting PANSS symptoms in schizophrenia spectrum disorders using speech only: an international, multi-centre, computational secondary analysis across multiple languages. eClinicalMedicine, 101, Article 104265. https://doi.org/10.1016/j.eclinm.2026.104265

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104265

Keywords: schizophrenia, speech analysis, machine learning, PANSS, multilingual, psychiatry, digital biomarkers, relapse prediction, self-supervised embeddings, mental health, Predicting, symptoms