ORBIT: New AI detects Alzheimer's across languages without prior training
Zero-shot cross-lingual Alzheimer detection using speech and text fusion breaks language barriers.
A new framework called ORBIT (cross-attentive fusion with multi-tap language adversaries and complementary spherical-hyperbolic geometric learning) tackles a critical challenge: detecting Alzheimer's disease from speech in languages never seen during training. Traditional AI models for speech-based Alzheimer's detection require labeled data in the target language, limiting scalability. The team behind ORBIT hypothesized that combining multilingual speech and text pretrained models through cross-attentive fusion would capture complementary acoustic and linguistic markers of cognitive impairment, while adversarial learning suppresses language-specific confounds.
Empirical results validate this approach: ORBIT consistently outperforms unimodal baselines and simple concatenation-based fusion on zero-shot cross-lingual evaluations. The framework also employs consensus clustering to further refine geometric representations. Accepted to INTERSPEECH 2026, ORBIT marks a significant step toward deployable, language-agnostic screening tools for Alzheimer's disease, potentially enabling broader access to early detection across global populations.
- ORBIT fuses multilingual speech and text features via cross-attentive fusion and multi-tap adversarial learning to remove language-specific biases.
- Uses complementary spherical and hyperbolic geometric learning with consensus clustering to capture cognitive impairment markers.
- Achieves strongest zero-shot cross-lingual performance, outperforming both unimodal models and simple fusion baselines.
Why It Matters
Enables Alzheimer's screening in any language without prior training, scaling early detection worldwide.