News / Detailed article

Can Speech Alone Reveal Psychosis Symptoms? TRUSTING Study Shows Promising Results

10 February, 2026

The EU-funded TRUSTING Project has achieved a major research milestone in Work Package 2 (WP2). Showing that speech recordings alone can help predict symptom severity in people living with psychosis, without analysing what is actually being said.

Predicting Psychosis Symptoms Using Speech Audio Only

In a large international, multi-centre retrospective study covering ten languages. Researchers examined whether symptoms measured by the Positive and Negative Syndrome Scale (PANSS) could be predicted using speech audio only, without relying on transcripts or spoken content.

Instead of analysing words, researchers extracted a wide range of acoustic speech features, including rhythm, pitch variation and voice dynamics. The results demonstrate that individual PANSS symptom scores can be predicted with reasonable accuracy, and importantly, the models worked consistently across different languages.

Largest Multilingual Speech Dataset in Psychosis Research

To the researchers’ knowledge, the study is based on the largest and most linguistically diverse speech dataset in psychosis research to date, representing a major step forward for global and multilingual mental health research.

Towards Low-Burden Digital Monitoring in Mental Health Care

The findings provide strong proof of concept that psychosis symptom monitoring may be possible without analysing speech content, opening the door to low-burden and scalable monitoring tools.

Such tools could help clinicians track symptom changes over time and potentially identify relapse risks earlier, supporting faster intervention and improved care.

Speech-based monitoring approaches could ultimately enable real-time, accessible mental health monitoring, reduce clinical workload, and support more personalised care for people living with psychosis worldwide.

Supporting Trustworthy Digital Mental Health Innovation

This milestone brings the TRUSTING project one step closer to delivering trustworthy, data-driven digital tools for mental health services, while keeping patient burden low and ensuring solutions remain scalable across languages and healthcare systems.

The workflow of this study. We analysed speech samples from audio recordings of 453 patients with schizophrenia spectrum disorders, recruited across ten global sites. After preprocessing the recordings and splitting them into shorter segments, we extracted three types of speech features: acoustic-prosodic features, pretrained multilingual embeddings, and a combination of the two. The data was further split into 80% train, 10% test, and 10% validation datasets. We then compared 16 algorithms, including machine learning and deep learning, to predict eight relapse-related PANSS items, including three positive (P1, P2, P3), three negative (N1, N4, N6), and two general (G5, G9) items. Performance was assessed by root-mean-squared-error (RMSE) at both segment and participant (median aggregation) levels.

For more updates, insights, and news about TRUSTING’s progress, visit the TRUSTING Project website or follow us on social media:

Funded by the European Union. Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or of the European Health and Digital Executive Agency (HaDEA). Neither the European Union nor the granting authority can be held responsible for them.