Speech-based machine learning detects Alzheimer’s across English and Greek cohorts

Researchers have developed an acoustic-only machine-learning pipeline that detects Alzheimer’s disease from speech recordings, reporting mean classification accuracies of 90.4% on an English cohort, 89.5% on a Greek cohort, and 86.2% on a mixed cohort of mild-to-moderate dementia patients and healthy controls. The study, published in Multimedia Tools and Applications, addresses a recognised gap: most speech-based Alzheimer’s detection research has been built on English corpora, leaving cross-linguistic evidence scarce.

The motivation is practical. Neuroimaging and cerebrospinal-fluid biomarkers are accurate but costly and largely confined to specialist centres, while low-cost, non-invasive tools for case detection at the mild-to-moderate stage remain an urgent need. Speech offers an accessible alternative, but its usefulness across different languages had been unclear.

The pipeline extracts multidomain acoustic features spanning the time, frequency and spectral, cepstral, and silence domains. Rather than relying on opaque selection, the authors screened features with a non-parametric Mann–Whitney U-test and cross-checked the results under Benjamini–Hochberg false-discovery-rate control, aiming for statistical transparency. Five classifiers were then compared for the binary task of distinguishing healthy controls from Alzheimer’s patients.

Methodological rigour was a stated priority: the full pipeline was repeated over ten random seeds under leakage-controlled, subject-stratified five-fold cross-validation, with every metric computed at the subject level and supported by bootstrap confidence intervals and paired significance tests. A support vector machine achieved the highest mean accuracy on each cohort, but on the mixed cohort it was not significantly better than logistic regression, K-nearest neighbours, or random forest — a nuance the authors report explicitly rather than overstate.

The cross-corpus transfer experiment is the study’s central finding. A compact subset of spectral-envelope features transferred across the English and Greek datasets, a pattern consistent with a shared, language-agnostic articulatory component in Alzheimer’s-related speech change. However, the authors caution that because language and elicitation task are aliased across the two cohorts, this shared component cannot be attributed to language alone. The work suggests acoustic markers of the disease may partly transcend linguistic boundaries, but disentangling language from task design will require further study.

Subject of Research: Cross-linguistic acoustic machine learning for detecting Alzheimer's disease from speech

Article Title: Harnessing cross-linguistic speech analysis for Alzheimer’s disease detection

Article References: El Hallani, A., Chakhtouna, A., & Adib, A. (2026). Harnessing cross-linguistic speech analysis for Alzheimer’s disease detection. Multimedia Tools and Applications, 85(11), Article 820. https://doi.org/10.1007/s11042-026-21976-2

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21976-2

Keywords: Alzheimer's disease, speech analysis, machine learning, acoustic features, cross-linguistic, support vector machine, spectral features, cross-validation, dementia detection, biomarkers, Harnessing, speech