A Language-Independent Speech-Based Classifier For Early Detection of Alzheimer’s Disease

The burden of Alzheimer’s disease demands accessible early detection tools. Speech analysis offers an accessible method, but current speech-based classifiers are monolingual, yet over 8,000 languages lack AD speech datasets, leaving millions unserved. Amyloid-β protein accumulation in the brain disrupts the left inferior frontal gyrus, dorsomedial prefrontal cortex, and posterior superior temporal sulcus, producing flattened prosody, irregular pauses, and slowed articulation. We propose the first truly language-independent acoustic biomarker classifier by extracting prosodic entropy, pause variability, articulation rate, and other features from speech data. Trained solely on English, our model generalizes zero-shot to diverse languages, enabling accessible global early AD screening.