New research, co-authored by Canary Speech and Japan’s National Cerebral and Cardiovascular Center (NCVC) scientists and published in The Lancet Regional Health – Western Pacific, demonstrates that voice alone can be a powerful tool for identifying mild cognitive impairment (MCI). Using short, three-minute unstructured conversations and machine learning models trained on voice biomarkers, the study achieved area under the curve (AUC) scores of 0.88 and 0.89, outperforming traditional models based only on age, sex, and education (AUCs 0.78–0.80). The difference was statistically significant (p < 0.0001).
The study analyzed 1,461 adults living independently in Japan, with 36% meeting criteria for cognitive impairment based on the Memory Performance Index. Researchers extracted 512-dimensional acoustic and prosodic features using Wav2Vec2, and trained both deep learning and XGBoost models. Importantly, the inclusion of voice biomarkers improved model performance by nearly 10 percentage points, highlighting voice as a clinically valuable, non-invasive indicator of early cognitive decline.
This is one of the largest and most rigorously validated studies to date showing that AI can detect cognitive impairment from voice alone, in under three minutes, without needing to analyze word content. As 92% of MCI cases currently go undiagnosed, the implications for population-scale screening are profound. Canary Speech is proud to have contributed to this groundbreaking research.
Interested in using voice as a powerful tool for early, accessible health detection? Email us at info@canaryspeech.com.
Reference: Eri Kiyoshige, Soshiro Ogata, Namhee Kwon, Yuriko Nakaoku, Chisato Hayashi, Nate Blaylock, Raymond Brueckner, Vinod Subramanian, Henry Joseph OConnell, Yusuke Yoshikawa, Kanako Teramoto, Kiyomasa Nakatsuka, Satoshi Saito, Masafumi Ihara, Misa Takegami, Kunihiro Nishimura, Developing and testing AI-based voice biomarker models to detect cognitive impairment among community dwelling adults: a cross-sectional study in Japan, The Lancet Regional Health – Western Pacific, Volume 59, 2025, 101598, ISSN 2666-6065, https://doi.org/10.1016/j.lanwpc.2025.101598.
