Researchers from Canary Speech, in collaboration with Asociación Grupo Ermita Alzheimer de Guatemala, evaluated vocal biomarkers for Alzheimer’s disease detection using a new Latin American Spanish dataset collected in Guatemala.
The study compared multiple machine-learning approaches using audio-only data and found that the best-performing model achieved an unweighted average recall of 0.82, with 0.81 sensitivity and 0.83 specificity. The researchers also evaluated prediction consistency under simulated real-world recording conditions, including background noise, changes in microphone characteristics, room acoustics, and loudness. Across these conditions, the model produced the same predicted classification in 90–98% of cases.
The findings contribute to research on vocal biomarkers beyond English-language datasets and provide additional insight into the reliability of audio-based approaches under varied recording conditions.
Reference: Raymond Brueckner, Namhee Kwon, Vinod Subramanian, Rachel Noack, Nate Blaylock, Henry. “Detection of Early Alzheimer’s Disease Using Vocal Biomarkers in Latin American Spanish”, Proceedings of IberSPEECH 2026, November, 2026.
